The analysis of rural tourism image optimization under the internet of things and deep learning Authors: Xinghua Wang Published: 2024 ABSTRACT This study aims to utilize deep learning technology to optimize rural tourism image, enhance visitor experience, and promote sustainable development. By deploying sensors for real-time monitoring of the environment and visitor flow in rural scenic areas, combined with a Dense Convolutional Neural Network (DenseNet), automatic identification and analysis of rural landscapes are achieved. Using rural tourism along the Yellow River as a case study, this study constructs a tourism image evaluation and optimization model based on big data. The results indicate that the model performs excellently in terms of accuracy and robustness, significantly improving the presentation of rural tourism images. The study shows that realism and service facilities have the greatest impact on rural tourism image, underscoring the value of technological means in optimizing the rural tourism image. Introduction In contemporary society, characterized by the relentless progression of urbanization and the rapid pace of modern life, the appeal of rural tourism as a means to seek tranquility, connection with nature, and immersion in tradition has gained increasing attention and affection1–3. Beyond merely fulfilling individuals’ yearning for natural surroundings, cultural immersion, and a traditional lifestyle, rural tourism plays a pivotal role in fostering economic development and generating employment opportunities in rural areas4,5. As the tourism industry continues to expand, the optimization of the rural tourism image emerges as a crucial factor in both attracting visitors and enhancing their overall experience. Tourists’ expectations for rural scenic spots extend beyond mere appreciation of natural landscapes; there is a growing emphasis on improving service facilities and providing enriched cultural experiences. Consequently, addressing the comprehensive optimization of the rural tourism image to enhance its allure and competitiveness has become an urgent and imperative challenge. The concept of integrated development has been gradually introduced into rural tourism, trying to realize the optimization and sustainable development of rural tourism through scientific and technological means based on maintaining rural characteristics6,7. As a bridge connecting the physical and digital worlds, the Internet of Things (IoT) technology offers new opportunities for rural tourism8. As a vital branch in the field of artificial intelligence, deep learning (DL) can extract and learn features from massive data, providing innovative means for image perception and evaluation of rural tourism9–11. The integration of IoT and DL technologies into rural tourism holds the promise of enhancing its attractiveness and sustainability across multiple dimensions, thereby improving tourist satisfaction and participation12,13. In addition, these technologies provide data and technical support for the integrated development of rural tourism from environmental perception, resource monitoring, and intelligent management14. In rural tourism settings, IoT technology enhances intelligent operation by deploying sensors, monitoring equipment, and network systems to achieve real-time monitoring and management of the environment, visitor behavior, and resource utilization. For instance, IoT devices can monitor environmental parameters such as air quality, temperature, and humidity within scenic areas and issue timely warnings about potential environmental issues, providing data support for resource scheduling and management. Regarding visitor experience, IoT technology offers personalized services through smart navigation, visitor flow monitoring, and real-time feedback. For example, visitors can receive location-based personalized recommendations via smart devices, providing information about nearby natural landscapes, cultural heritage, and entertainment facilities. IoT can also analyze real-time visitor data, such as crowd density and wait times, to dynamically adjust service strategies, avoiding overcrowding or resource waste, thus improving overall visitor satisfaction. DL technology improves the accuracy of image presentation and resource management in scenic areas. It does this by automatically analyzing and extracting features from large amounts of tourism images, videos, and other data. Convolutional neural networks (CNNs) in DL, especially deep architectures like DenseNet, can process many images taken by visitors. These images come from sources such as tourism websites and social media platforms. CNNs perform tasks like landscape recognition, sentiment analysis, and visual aesthetic evaluation. These analyses enable scenic area managers to identify popular attractions and pinpoint shortcomings in services or facilities, allowing targeted optimization. Additionally, DL technology can analyze visitor behavior patterns and preferences to automatically generate personalized travel recommendations, enhancing the customization of the tourist experience. By integrating real-time environmental data collected by IoT, DL models can more promptly and accurately respond to visitor needs. For example, certain scenic areas can automatically recommend different tour routes and activities based on weather changes or visitor preferences, improving visitor satisfaction and engagement. In terms of resource management, the combination of IoT and DL technology facilitates intelligent scenic area management. By utilizing real-time environmental and visitor flow data collected through IoT, DL models can predict visitor movement trends and peak resource demand, helping managers optimize resource allocation and avoid overloading facilities or wasting resources. This not only improves operational efficiency but also reduces environmental impact, contributing to the sustainable development of rural tourism. However, rural tourism also faces some challenges in the context of integrated development. On the one hand, the infrastructure and services in rural areas are relatively lagging behind, and the tourist experience may be affected15,16. On the other hand, protecting rural areas’ natural environment and cultural traditions and achieving sustainable development has also become a difficult problem for rural tourism17,18. The primary objectives of this study are as follows: (1) to enhance the representation of rural tourism images through DL technology, thereby improving landscape recognition and attracting a greater number of tourists; (2) to integrate Internet of Things (IoT) and DL technologies to enrich the tourist experience by offering real-time environmental monitoring and interactive services; (3) to promote the sustainable development of rural tourism by balancing ecological preservation with economic growth; and (4) to support data-driven decision-making for improved management and sustainability of rural scenic areas. This study seeks to provide both theoretical and empirical foundations for optimizing rural tourism images, ultimately contributing to the sector’s sustainable development. To achieve these aims, the study proposes a comprehensive approach, incorporating a rural tourism image perception model based on the Dense Convolutional Neural Network (DenseNet), alongside a rural tourism image evaluation and optimization model utilizing big data. This framework introduces innovative concepts and methodologies for advancing rural tourism development19,20. The study analyzes how IoT and DL technologies are applied in rural tourism to enhance the efficiency of data processing and analysis, optimizing various aspects of visitor experience and resource management. IoT technology not only provides data support for environmental monitoring and resource allocation in tourism areas, but also, when combined with DL technology, improves the accuracy of tourism image recognition. This, in turn, enhances the overall competitiveness and sustainable development of rural tourism. The study is divided into five sections. The first section, the introduction, presents the research background, motivation, and objectives. The second section is a literature review that examines the current state of research on rural tourism image optimization both domestically and internationally, analyzes the limitations of existing studies, and outlines the potential applications of IoT and DL technologies in this field. The third section introduces the research model, which includes the construction of a rural tourism image perception model based on DenseNet and the development of an evaluation and optimization model for rural tourism images using big data. The fourth section details the experimental design and performance evaluation, describing the data collection process, the experimental environment setup, parameter settings, and the performance metrics used to validate the model’s effectiveness. The fifth section provides conclusions and discussions, summarizing the key findings and contributions of the research, and proposing directions for future research. Literature review Internationally, there is a growing focus on research related to the optimization of rural tourism images through the integration of IoT and DL technologies within the context of integrated development. Notably, investigations into these technologies have yielded extensive research outcomes, significantly contributing to the enhancement and refinement of rural tourism images. Sarker focused on realizing environmental monitoring of rural scenic spots and statistics of tourist flow by deploying sensors, monitoring equipment, and other means to provide real-time data support for rural tourism21. Merenda emphasized using DL technology to improve the accuracy of rural landscape image recognition and classification, and exploring tourists’ emotions and opinions on social media through sentiment analysis22. In terms of image optimization of rural tourism, Yang focused on the publicity and promotion of rural tourism, and discussed how to enhance the visibility and attractiveness of rural tourism through new media and digital means23. With the proposal of the concept of urban-rural integrated development, Li paid attention to the role of rural tourism in promoting the rural economy and protecting rural culture, and began to investigate how to balance ecological protection and economic growth in rural tourism from the perspective of sustainability24. Luo tried to apply IoT and DL technologies to image optimization in some rural tourism scenic spots. For example, some scenic spots introduced intelligent navigation systems to provide tourists richer interactive experiences25.Kua explored how to realize a smart tourism experience through IoT and DL technologies in the domain of rural tourism, using sensors, mobile applications, and other technologies to provide personalized guides and recommendations for tourists and enhance their interactive experience26. Wang (2023) was committed to developing algorithms that automatically identify and classify rural landscapes, thus offering tourists more detailed and accurate information, increasing their interest and satisfaction27. In rural tourism image optimization, Zhu focused on how to employ DL technology to analyze the comments and feedback of tourists on social media. They utilized methods such as sentiment analysis to understand tourists’ perceptions of rural tourism to optimize tourism products and services28. Liao examined the combination of IoT and DL technologies with virtual reality (VR) and augmented reality (AR) to create a richer rural tourism experience, presenting visitors with virtual landscapes, historical and cultural reproductions, and others29. Additionally, significant progress has been made in applying IoT and DL technologies to rural tourism in Western countries. In Europe, particularly in Italy, Moise et al. explored how IoT technology enhanced the sustainability of rural tourism and improved the visitor experience through smart devices and sensors. The result showed that by collecting real-time data, rural attractions could optimize resource allocation, reduce environmental pressure, and increase visitor satisfaction. Furthermore, the research highlighted that the introduction of smart parking, smart lighting, and other technologies not only enhanced visitor convenience but also created new revenue streams for local economies30. In the United States, Wang investigated the application of DL technology in the rural tourism market, emphasizing its potential for predicting visitor behavior and providing personalized recommendations. By analyzing social media data and visitor feedback, the study demonstrated that DL models could effectively identify visitor preferences, enabling the provision of more accurate tourism products and services, thereby improving visitor satisfaction and loyalty31. In Canada, Green et al. (2021) conducted research on the application of IoT technology in environmental monitoring within rural tourism, particularly in ecotourism. They showed that by deploying sensor networks, real-time monitoring of environmental changes and assessments of ecosystem health could be conducted. This not only helped protect natural resources but also provided visitors with an enriched educational experience, helping them better understand the importance of ecological conservation32. The introduction of IoT and DL technologies has created new economic opportunities for rural tourism. Li pointed out that IoT and DL technologies could enhance the operational efficiency of scenic areas, reduce management costs, and attract more tourists through targeted marketing efforts33. This not only stimulates local economic growth but also creates job opportunities for local communities. For example, the introduction of smart tour guide systems enables rural tourist destinations to offer more diverse services, thereby increasing visitor spending. On a societal level, the application of IoT and DL technologies improves the overall visitor experience and strengthens interaction between visitors and local communities. Zhang noted that by using social media analytics, rural tourism managers could better understand tourists’ needs and preferences, allowing them to adjust services and activities to improve visitor satisfaction and engagement34. This interaction also deepens visitors’ understanding and appreciation of local culture, contributing to the promotion and preservation of community culture. However, the application of these technologies also presents potential environmental challenges. Chen warned that an over-reliance on technology could lead to the overexploitation of resources, particularly in environmentally sensitive areas35. While the use of smart devices can increase resource use efficiency, if not properly managed, it may exert pressure on the local ecosystem. Therefore, researchers recommend that rural tourism development be combined with ecological conservation, implementing effective management strategies to ensure sustainable development. The application of IoT and DL technologies in rural tourism in Western countries demonstrates their potential in enhancing visitor experience, promoting economic growth, and protecting the environment. However, differences in local characteristics and socio-cultural contexts lead to varying effects of technology application across different regions. In summary, researchers in China and other countries conducted extensive research in this field. They actively explored innovation in intelligent tourism experience, landscape identification, social media analysis, sustainable development, and other aspects, providing a wealth of ideas and methods for rural tourism development. However, previous studies focused on different aspects of rural tourism image optimization, such as environmental monitoring, image recognition, promotion, economic promotion, etc. Meanwhile, there was a lack of a comprehensive framework that fully integrated IoT and DL technologies, and the impact of big data on decision-making and evaluation was not fully considered. Furthermore, there was not a thorough examination of the total picture, and the emphasis on the visitor experience was very dispersed. This study’s goal is to address these shortcomings by integrating IoT and DL technologies, building on prior research, and developing a comprehensive model for optimizing rural tourism images. The ultimate goal is to enhance visitor experiences and encourage the sustainable growth of this industry. Model and methodological approach First, real-time monitoring of environmental conditions and tourist flows is achieved by deploying IoT technology in rural scenic areas, including sensors and monitoring equipment. The application of this IoT technology provides accurate real-time data for subsequent research, providing important support for decision-making and sustainable development. In terms of DL technology, DenseNet is chosen as the basic model for image recognition to automatically identify and analyze rural landscapes. The key to this step is to conduct pre-processing and feature extraction after acquiring image data to ensure that the data meets the requirements of the DenseNet model. Meanwhile, the visual feature data of Mafengwo.com photos are extracted with the help of this model, which lays a foundation for the subsequent image evaluation and optimization. Based on the above work, a big data-driven rural tourism image evaluation and optimization model is established. The actual data of villages along the Yellow River is used for testing. During the model testing and performance evaluation phase, careful attention is given to several crucial indicators, including the accuracy of image recognition, data loss rate, and the correlation between visual aesthetic quality and location. This meticulous approach aims to thoroughly assess the model’s practical applicability. Through a rigorous analysis of the test results, conclusive insights can be gained regarding the authenticity of the image and the influence of service facilities on the rural tourism image. Second, this research method is compared with previous studies by means of comparison and discussion, highlighting its innovation and comprehensiveness in the field of rural tourism. Lastly, combined with the experimental results, the conclusions on the optimization of rural tourism image are put forward, and the possible improvement and further research direction are looked forward. Through this comprehensive research method, this study aims to offer strong theoretical and empirical support for rural tourism image optimization and sustainable development. In this study, the DenseNet model is adopted and applied to the field of rural tourism image perception, which is a pioneering attempt in this field and improves the accuracy of landscape identification and personalized recommendation. Secondly, this study carries out detailed preprocessing and feature extraction on the large-scale image data from Mafengwo.com official travel websites to ensure the reliability and effectiveness of model training. In addition, it verifies the practicability and reliability of the model through comprehensive performance evaluation, including accuracy, data loss rate, recall rate and other indicators. Finally, it introduces the experimental results in detail, highlighting the role and significance of the model in rural tourism image optimization, including the experimental results of improving accuracy, controlling data loss rate, and the correlation between image characteristics and geographical location. Therefore, this study is innovative in the application of DenseNet model and has made positive contributions to the field of rural tourism image optimization. Models Enhancing visitor experience through IoT technology The application of IoT technology in rural tourism destinations is primarily reflected in real-time monitoring of the environment and visitor flow. By deploying a multitude of sensors throughout the site, data on air quality, temperature, humidity, noise levels, and other environmental factors can be collected. This data, combined with visitor flow analysis, allows for dynamic adjustments to operational strategies, thereby enhancing the visitor experience. For instance, air quality sensors can detect changes in air quality within the area. When the concentration of air pollutants rises, the system can promptly alert site managers to take necessary measures or notify visitors through a mobile app, suggesting they temporarily avoid that area. This approach ensures that visitors enjoy a more comfortable travel experience while safeguarding their health and safety. Simultaneously, the site can utilize real-time monitoring data to effectively manage visitor flow, preventing overcrowding and improving the quality of the visit. In terms of monitoring visitor flow, sensors can record the entry and movement paths of visitors, allowing for personnel diversion through an intelligent scheduling system. Site managers can adjust visitor directions based on real-time data, preventing excessive concentration of visitors in specific areas, thereby reducing wait times and crowding, which enhances visitor satisfaction. Additionally, monitoring data can be used to predict peak periods, providing advance warnings and facilitating appropriate reception preparations. IoT technology not only enhances the visitor experience but also plays a crucial role in promoting the sustainable development of rural tourism. By monitoring environmental factors with sensors, site managers can dynamically assess the health of the ecosystem, enabling the formulation of more scientifically informed management and conservation measures. For example, when water sources are contaminated or vegetation is damaged, sensors can issue timely alerts, allowing managers to swiftly implement restoration measures to mitigate the negative impacts of tourism activities on the natural environment. Furthermore, environmental data obtained through sensors can support long-term planning and development in rural areas. By analyzing visitor behavior and environmental feedback, destinations can gradually optimize infrastructure development, avoiding overexploitation that could harm natural resources, thereby achieving a balanced development of economic, social, and ecological factors. This data-driven management approach ensures the sustainable utilization of rural tourism resources and aids in achieving the dual goals of economic revitalization and ecological protection for local communities. A rural tourism image perception model integrating densenet DenseNet is a kind of DL neural network structure with the characteristic of dense connection. Compared with traditional CNNs, DenseNet al.lows the output of each layer to be directly transmitted to all subsequent layers by introducing dense connections into the Network, thus achieving more effective information transmission and gradient flow. This structure helps alleviate the problem of gradient vanishing, enhances feature reuse, and improves the performance and generalization ability of the network. The core idea of DenseNet is to achieve efficient feature propagation through dense connections. In each Dense Block, the output of a previous layer is not only fed into the current layer but also into all subsequent layers. This type of connectivity significantly enhances feature utilization and reduces computational complexity during training, thereby improving the model’s recognition and classification accuracy. In practice, the design of DenseNet involves several key hyperparameters, including the number of network layers, growth rate, and compression factor. Choosing the number of layers requires balancing model complexity and performance. Generally, increasing the number of layers can enhance the model’s representational power, but it also increases the number of parameters and computational complexity. The growth rate refers to the increment in the dimensionality of the feature maps output by each Dense Block. A higher growth rate can improve the network’s nonlinearity and feature extraction capabilities, but it also increases the number of parameters. The compression factor is used to control the reduction in the dimensionality of the feature maps, helping to decrease the number of parameters and computational complexity. A smaller compression factor may lead to overfitting, while a larger compression factor could negatively affect the model’s representational capacity. DenseNet has achieved excellent results in many competitions in image recognition tasks. So, this study chooses it as the basic model. The rural tourism image perception model, based on DenseNet and rooted in DL technology, strives to achieve effective perception and feature extraction of rural tourism landscape images36–38. The design of this model is inspired by the concepts of traditional CNNs and dense connectivity, aiming to overcome the problem of gradient vanishing in traditional CNNs and improve the accuracy of image recognition and classification39–41. The structure of DenseNet is presented in Fig. 1: Fig. 1DenseNet structure. DenseNet structure. In Fig. 1, DenseNet connects the previous layer’s output with the input of all subsequent layers by introducing a dense layer, so that information can be more fully transmitted in the network42. This dense connection effectively addresses the problem that information gradually decreases when transmitted layer by layer in the CNN. Hence, the network can better capture the details and features of rural landscape images13,43. The specific processing process of the model is as follows: Step 1: Data preprocessing;Rural tourism landscape image datasets are preprocessed, including image scaling, normalization, data enhancement, etc., to ensure data quality consistency44,45.Step 2: Data input;After processing, image data is input.Step 3: Densely connected block passing;In each densely connected block, the output of each convolutional layer is connected with the output of all previous layers. The gradient is maintained during data transfer, and the data characteristics are maintained46;Step 4: Extraction of data features;The feature map’s size is decreased, reducing the amount of calculation. The transition layer can include a convolutional layer and a pooling layer, which gradually reduces the size of the feature map to achieve the gradual extraction of information47–49.Step 5: Establishment of image perception mechanism;The image perception mechanism of rural tourism is plotted in Fig. 250: Step 1: Data preprocessing;Rural tourism landscape image datasets are preprocessed, including image scaling, normalization, data enhancement, etc., to ensure data quality consistency44,45. Step 1: Data preprocessing; Rural tourism landscape image datasets are preprocessed, including image scaling, normalization, data enhancement, etc., to ensure data quality consistency44,45. Step 2: Data input;After processing, image data is input. Step 2: Data input; After processing, image data is input. Step 3: Densely connected block passing;In each densely connected block, the output of each convolutional layer is connected with the output of all previous layers. The gradient is maintained during data transfer, and the data characteristics are maintained46; Step 3: Densely connected block passing; In each densely connected block, the output of each convolutional layer is connected with the output of all previous layers. The gradient is maintained during data transfer, and the data characteristics are maintained46; Step 4: Extraction of data features;The feature map’s size is decreased, reducing the amount of calculation. The transition layer can include a convolutional layer and a pooling layer, which gradually reduces the size of the feature map to achieve the gradual extraction of information47–49. Step 4: Extraction of data features; The feature map’s size is decreased, reducing the amount of calculation. The transition layer can include a convolutional layer and a pooling layer, which gradually reduces the size of the feature map to achieve the gradual extraction of information47–49. Step 5: Establishment of image perception mechanism; Step 5: Establishment of image perception mechanism; The image perception mechanism of rural tourism is plotted in Fig. 250: The image perception mechanism of rural tourism is plotted in Fig. 250: Fig. 2The image perception mechanism of rural tourism. The image perception mechanism of rural tourism. Figure 2 describes the image perception of rural tourism from three aspects: perception dimension, perception process, and dynamic restrictions. In Fig. 2, “primitive cognition” and “spatial perception” are two important parts of the image perception mechanism, which represents different concepts and meanings respectively. Firstly, “original cognition” refers to tourists’ first impression and cognition of scenic spots, that is, tourists’ intuitive feelings and cognition when they first come into contact with scenic spots. At this stage, tourists may have a preliminary understanding and cognition of the overall atmosphere and landscape characteristics of the scenic spots, and form a preliminary perception impression. In contrast, “spatial perception” focuses more on tourists’ perception and understanding of the spatial layout and location of scenic spots. Through spatial perception, tourists can have a deeper understanding of the geographical location, layout structure and the relationship with the surrounding environment of scenic spots, thus forming a cognition of the spatial characteristics of scenic spots. The cognition at this stage usually needs more observation and experience, and tourists will gradually establish a comprehensive cognition of the spatial structure of scenic spots. Therefore, “primitive cognition” and “spatial perception” play different but complementary roles in the process of tourists’ cognition of scenic spots. There is a close relationship between “primitive cognition” and “spatial perception”. Tourists will be influenced by their original cognition while they perceive space. For example, tourists’ initial impression of scenic spots may affect their perception and understanding of scenic spot space. In the process of spatial perception, tourists’ in-depth understanding of the spatial layout and characteristics of scenic spots will further affect the formation and adjustment of their original cognition. Therefore, “primitive cognition” and “spatial perception” interact and influence each other, and together they form a comprehensive cognitive system for tourists to scenic spots. In the process of image-building and marketing of tourist destinations, understanding and grasping the relationship between them can better guide tourists’ perception and cognition, and enhance the attraction and sustainable development ability of scenic spots. Firstly, in terms of perception dimension, image perception includes space, time, psychology and environment, spatial perception refers to tourists’ perception and understanding of the spatial layout and location of scenic spots. Time perception involves tourists’ perception of the time arrangement and historical evolution of tourism activities. Psychological perception includes tourists’ feelings, emotions and attitudes in the process of traveling. Environmental perception refers to tourists’ perception of the natural, cultural and social aspects of the tourism environment. Secondly, in the process of perception, Fig. 2 describes several key steps involved in the process of image perception. These steps include the generation of perception, original cognition, compound cognition, promotion cognition and restriction cognition. The generation of perception refers to tourists’ first impression and cognition of scenic spots. The original cognition involves the initial cognition and understanding of the characteristics of scenic spots. Compound cognition is to combine and integrate the characteristics of different scenic spots on the basis of original cognition. Promoting cognition refers to promoting tourists’ in-depth understanding and cognition of scenic spots through various means. Restricted cognition is all kinds of restrictions and obstacles that tourists encounter in the cognitive process. Finally, in the aspect of dynamic restriction, Fig. 2 emphasizes the dynamic restriction factors such as feature reservation, feature endowing, feature spreading and feature management. Feature preservation refers to the preservation and dissemination of scenic features and images in the process of image perception. Giving features refers to identifying, understanding and giving meaning to the features of scenic spots. The spread of characteristics refers to spreading the characteristics of scenic spots to more tourists through various media and channels. Feature management refers to the effective management and maintenance of scenic features to ensure the sustainability and stability of the image of scenic spots. To sum up, Fig. 2 comprehensively and systematically shows the key contents of rural tourism image perception, which provides important theoretical and practical guidance for further research and optimization of rural tourism. Step 6: Result output;The global average pooling layer is added to reduce the dimensionality of each channel in the feature map to a scalar, which is used for the output of the final classification or regression task. Step 6: Result output;The global average pooling layer is added to reduce the dimensionality of each channel in the feature map to a scalar, which is used for the output of the final classification or regression task. Step 6: Result output; The global average pooling layer is added to reduce the dimensionality of each channel in the feature map to a scalar, which is used for the output of the final classification or regression task. Before the optimization scheme of DenseNet in rural tourism image processing, firstly, the network structure of DenseNet consists of several Dense Block, each of which consists of several convolution layers. When choosing the number of network layers, it is necessary to weigh the complexity and performance of the network. In general, increasing the number of network layers can improve the expressive ability of the model, but it will also increase the number of parameters and computational complexity. When determining the number of network layers, people need to consider the size of training data, the limitation of computing resources and the performance requirements of the model. Secondly, the Growth Rate in DenseNet refers to the dimension increase of the output feature map in each dense block. The choice of growth rate will affect the feature extraction ability and the number of parameters of the network. A larger growth rate can increase the nonlinearity of the network and improve the ability of feature extraction, but it will also increase the number of parameters and computational complexity. Therefore, when choosing the growth rate, it is needed to weigh the complexity of the dataset and the performance requirements of the model. In addition, the Compression Factor in DenseNet is used to control the dimension compression ratio of feature map to reduce the number of parameters and computational complexity of the model. A smaller compression factor will retain more feature information, but it may increase the risk of over-fitting of the model. The larger compression factor will reduce the number of parameters, but it may affect the expression ability of the model. Therefore, the performance and generalization ability of the model need to be considered comprehensively when choosing the compression factor. The possible limitations of current parameter setting include: too large or too small parameter setting may lead to over-fitting or under-fitting of the model. Too deep network layers may lead to gradient disappearance or explosion, making training difficult. The choice of growth rate and compression factor may affect the feature extraction ability and generalization ability of the model In view of the above limitations, the following improvement methods can be proposed: improving the model performance by adjusting the network structure or parameter optimization algorithm, such as using adaptive growth rate and compression factor. The attention mechanism is introduced to improve the model’s attention to key features, thus improving the model’s performance and generalization ability. Integrating other DL models further improves the accuracy and generalization ability of the model. Based on the existing research or experimental results, the optimization scheme of DenseNet in rural tourism image processing is as follows: introducing multi-scale attention mechanism to improve the model’s perception ability of different scale features to better capture the details and characteristics of rural landscape. Other DL models are integrated to improve the generalization ability and adaptability of the model. By adding data enhancement technology and model distillation, the performance and robustness of the model are further improved, so that it is better suitable for different rural tourism scenarios. In addition to DenseNet, this study also considers several other advanced DL models to provide a comprehensive evaluation and comparison of image recognition performance in rural tourism contexts. These models include Residual Networks (ResNet), Visual Geometry Group (VGG), EfficientNet, and Vision Transformer (ViT). Shafiq et al. (2022) pointed out that ResNet effectively addressed the vanishing gradient problem in deep network training by introducing residual connections (skip connections). This innovation made deeper network architectures possible while maintaining training stability. Research showed that ResNet performed exceptionally well in image classification tasks, especially when dealing with complex datasets, as it reached hundreds of layers while maintaining good performance. Therefore, ResNet served as a benchmark for DenseNet, helping analyze the impact of different network depths on image recognition capabilities51. Ya et al. noted that the VGG model was renowned for its depth and structural simplicity. VGG significantly enhanced feature extraction capabilities by using convolutional kernels of the same size and increasing network depth. Studies showed that VGG performed well in image classification tasks, particularly on smaller datasets. VGG served as another comparison model for DenseNet, especially in terms of its feature extraction efficiency and computational complexity52. Wang et al. highlighted that EfficientNet was an efficient model that achieved a remarkable balance between model size, computational cost, and accuracy through a compound scaling method. The design of this model took into account not only depth but also width and resolution, leading to state-of-the-art performance on multiple benchmark datasets. By comparing DenseNet with EfficientNet, the experiment further explored the trade-off between resource usage and accuracy53. Zhang et al. pointed out that the ViT model, which used a self-attention mechanism, altered the traditional CNNs structure, showing impressive performance in image classification tasks. ViT excelled in handling large-scale datasets, particularly when processing high-resolution images. The inclusion of ViT helped assess the impact of different architectures (convolutional vs. transformer) on image recognition54. Rural tourism image evaluation and optimization model based on big data This study chooses big data to establish the rural tourism image evaluation and optimization model. Big data has rich information and can fully reflect the multi-dimensional characteristics of rural tourism. Through the analysis of large-scale tourism images and relevant statistical data, the research can more comprehensively understand rural tourism’s visual characteristics and statistical information, providing a more accurate and comprehensive basis for image evaluation and optimization. “The rural tourism image evaluation and optimization model based on big data” is a method using big data analysis technology, aiming to evaluate and optimize the rural tourism image by collecting, analyzing, and mining large-scale rural tourism data, to improve the experience and attraction of tourists55,56. The Analytic Hierarchy Process (AHP) is a multi-criteria decision analysis method that is used to evaluate and rank the importance of different factors. In the evaluation of rural tourism image, AHP can help establish the hierarchical structure among various factors, and finally find the key factors that have the greatest impact on rural tourism image through the weight distribution of each level. AHP was selected because it may give an organized and methodical approach to synthesize and analyze a variety of complicated aspects and offer scientific backing for decision-making. As a result, this study uses the AHP to create this model. The steps of the AHP include: first, constructing a multi-level judgment matrix to define the main and sub-criteria, and involving experts to develop a comparison matrix to determine the weight distribution of each criterion. Through data collection and weight calculation, an evaluation index system is established, followed by a consistency check to ensure the reliability of the model. Ultimately, based on the scoring results, AHP proposes an optimization plan for rural tourism images. Compared to methods such as Multi-Attribute Utility Theory (MAUT) and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), AHP has a more systematic hierarchical structure, effectively addressing complex decision-making problems and providing more scientifically sound decision support. The steps involved are as follows: Step 1: Establishment of multi-level discriminant layer;The setting of the primary index is shown in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A=\left\{ {{A_1},{A_2}, \cdots ,{A_n}} \right\}$$\end{document}A represents the overall indicator; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{A}}_{\text{i}}$$\end{document} means the primary index. The secondary index is expressed in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A_i}=\left\{ {{a_1},{a_2}, \cdots ,{a_n}} \right\}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{\text{i}}$$\end{document} refers to the secondary index.Step 2: Experts are invited to cooperate in constructing a comparison matrix, and the evaluation set is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U=\left\{ {{u_1},{u_2}, \cdots ,{u_n}} \right\}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{\text{i}}$$\end{document} indicates the evaluation criteria.Step 3: Determination of the weight allocation of indicators. Primary and secondary indexes are introduced, the weights for each layer of indexes are calculated, and the weights are greater than 0.Step 4: Establishment of an evaluation index system;Step 5: Data are collected, and indicator weights are calculated;The expression of the data normalization processing reads:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_{ij}}=\frac{{{r_{ij}} \cdot }}{{\mathop \sum \nolimits_{{i=1}}^{n} {r_{ij}}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{w}}_{ij}$$\end{document} represents the weight, and the feature vector is calculated as Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_i}=\mathop \sum \limits_{{j=1}}^{n} {w_{ij}}$$\end{document}The weight vector of each factor is obtained by normalization.Step 6: Consistency test;This is defined as Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U*W={\lambda _{max}}*w$$\end{document}U stands for the contrast matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{max}$$\end{document} represents the maximum eigenvalue, and the calculation is:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda _{max}}=\frac{1}{n}\mathop \sum \limits_{{i=1}}^{n} \frac{{\left( {U{W_i}} \right)}}{{{w_i}}}$$\end{document}n refers to the order of the matrix. Random consistency is calculated by Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CR=\frac{{CI}}{{RI}}$$\end{document}RI represents a portability index. The judgment matrix is considered consistent if the CI is less than 0.1.Step 7: The results are analyzed, and an optimization scheme is proposed. The proposed rural tourism image optimization model is depicted in Fig. 3. Step 1: Establishment of multi-level discriminant layer;The setting of the primary index is shown in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A=\left\{ {{A_1},{A_2}, \cdots ,{A_n}} \right\}$$\end{document}A represents the overall indicator; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{A}}_{\text{i}}$$\end{document} means the primary index. The secondary index is expressed in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A_i}=\left\{ {{a_1},{a_2}, \cdots ,{a_n}} \right\}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{\text{i}}$$\end{document} refers to the secondary index. Step 1: Establishment of multi-level discriminant layer; The setting of the primary index is shown in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A=\left\{ {{A_1},{A_2}, \cdots ,{A_n}} \right\}$$\end{document} A represents the overall indicator; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{A}}_{\text{i}}$$\end{document} means the primary index. The secondary index is expressed in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A_i}=\left\{ {{a_1},{a_2}, \cdots ,{a_n}} \right\}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{a}_{\text{i}}$$\end{document} refers to the secondary index. Step 2: Experts are invited to cooperate in constructing a comparison matrix, and the evaluation set is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U=\left\{ {{u_1},{u_2}, \cdots ,{u_n}} \right\}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{\text{i}}$$\end{document} indicates the evaluation criteria. Step 2: Experts are invited to cooperate in constructing a comparison matrix, and the evaluation set is as follows:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U=\left\{ {{u_1},{u_2}, \cdots ,{u_n}} \right\}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{u}_{\text{i}}$$\end{document} indicates the evaluation criteria. Step 3: Determination of the weight allocation of indicators. Primary and secondary indexes are introduced, the weights for each layer of indexes are calculated, and the weights are greater than 0. Step 3: Determination of the weight allocation of indicators. Primary and secondary indexes are introduced, the weights for each layer of indexes are calculated, and the weights are greater than 0. Step 4: Establishment of an evaluation index system; Step 4: Establishment of an evaluation index system; Step 5: Data are collected, and indicator weights are calculated;The expression of the data normalization processing reads:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_{ij}}=\frac{{{r_{ij}} \cdot }}{{\mathop \sum \nolimits_{{i=1}}^{n} {r_{ij}}}}$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{w}}_{ij}$$\end{document} represents the weight, and the feature vector is calculated as Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_i}=\mathop \sum \limits_{{j=1}}^{n} {w_{ij}}$$\end{document}The weight vector of each factor is obtained by normalization. Step 5: Data are collected, and indicator weights are calculated; The expression of the data normalization processing reads:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_{ij}}=\frac{{{r_{ij}} \cdot }}{{\mathop \sum \nolimits_{{i=1}}^{n} {r_{ij}}}}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{w}}_{ij}$$\end{document} represents the weight, and the feature vector is calculated as Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w_i}=\mathop \sum \limits_{{j=1}}^{n} {w_{ij}}$$\end{document} The weight vector of each factor is obtained by normalization. Step 6: Consistency test;This is defined as Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U*W={\lambda _{max}}*w$$\end{document}U stands for the contrast matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{max}$$\end{document} represents the maximum eigenvalue, and the calculation is:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda _{max}}=\frac{1}{n}\mathop \sum \limits_{{i=1}}^{n} \frac{{\left( {U{W_i}} \right)}}{{{w_i}}}$$\end{document}n refers to the order of the matrix. Random consistency is calculated by Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CR=\frac{{CI}}{{RI}}$$\end{document}RI represents a portability index. The judgment matrix is considered consistent if the CI is less than 0.1. Step 6: Consistency test; This is defined as Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$U*W={\lambda _{max}}*w$$\end{document} U stands for the contrast matrix. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{max}$$\end{document} represents the maximum eigenvalue, and the calculation is:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\lambda _{max}}=\frac{1}{n}\mathop \sum \limits_{{i=1}}^{n} \frac{{\left( {U{W_i}} \right)}}{{{w_i}}}$$\end{document} n refers to the order of the matrix. Random consistency is calculated by Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CR=\frac{{CI}}{{RI}}$$\end{document} RI represents a portability index. The judgment matrix is considered consistent if the CI is less than 0.1. Step 7: The results are analyzed, and an optimization scheme is proposed. The proposed rural tourism image optimization model is depicted in Fig. 3. Step 7: The results are analyzed, and an optimization scheme is proposed. The proposed rural tourism image optimization model is depicted in Fig. 3. Fig. 3The structure of the rural tourism image optimization model. The structure of the rural tourism image optimization model. Figure 3 reveals that according to the evaluation results of the rural tourism image, the optimization of the rural tourism image needs to start from three perspectives: original image, overall image, and induced image, and establish a rural tourism image combining ecology, production, and life. The model covers the original image of rural tourism image, and regards it as a combination of functions, images, characteristics, space, ecology, media and society, emphasizing the diversity and comprehensiveness of the image. Through the concept of integrated development, optimization guidance and strategies are put forward to guide the rural tourism image to develop in a more perfect and attractive direction. In terms of optimization guidance, the importance of service facilities, authenticity and other factors is emphasized, which guides the promotion and improvement of rural tourism image. The optimization strategy pays more attention to the original characteristics of the image, focusing on the in-depth optimization and renewal of the rural tourism image from the aspects of ecology, space and society. Generally speaking, the model is based on the concept of integrated development, and guides the rural tourism image to develop and improve in a more perfect and attractive direction by providing comprehensive optimization guidance and strategies. To sum up, it is one of the key steps in this study to use DenseNet model to perceive and extract the characteristics of rural tourism landscape images. DenseNet is a kind of DL neural network structure, and its dense connection makes the output of each layer directly transmitted to all subsequent layers, which effectively solves the problem that information gradually decreases when it is transmitted layer by layer in traditional CNNs. This structure is helpful to extract the details and features of rural tourism landscape images, thus achieving more effective information transmission and gradient flow. In the process of model design, it uses DenseNet as the basic model, and realizes efficient perception and feature extraction of rural tourism landscape images through DL technology. The design of this model is inspired by the concepts of traditional CNN and dense connection, aiming at overcoming the problem of gradient disappearance in traditional CNN and improving the accuracy of image recognition and classification. In the process of perception and feature extraction, data preprocessing, including image scaling, normalization and data enhancement, is firstly carries out to ensure the consistency of data quality. Then, the processed image data is input into the DenseNet model. In each densely connected block, the output of each convolution layer is connected with the output of all previous layers, thus maintaining gradient transmission and maintaining data characteristics. Subsequently, the calculation amount is reduced through the transition layer where the size of the feature map is gradually reduced, and the information is gradually extracted. Finally, by adding a global average pooling layer, the dimension of each channel is reduced to a scalar, which is used for the output of the final classification or regression task. In the training process of DenseNet model, DL training parameters are used to train the model, such as learning rate and batch size to achieve the best convergence and generalization performance. In the testing stage, the relevant data obtained by using DenseNet model include image recognition results, confidence scores and so on, which provides a basis for the subsequent rural tourism image evaluation and optimization model. For example, the characteristic data obtained from the DenseNet model can be used to establish an image perception mechanism, and combined with the visual characteristics and statistical information of the image, it can provide data support for the comprehensive evaluation of rural tourism images. Next, the features extracted from the DenseNet model are graded by using the AHP to realize the comprehensive evaluation and optimization of rural tourism images. AHP is a multi-criteria decision analysis method, which can be used to evaluate and rank the importance of different factors. In the evaluation of rural tourism images, AHP can help to establish the hierarchical structure among various factors, and finally find the key factors that have the greatest influence on rural tourism images through the weight distribution of indicators at all levels. Firstly, a multi-level discrimination layer is established, the main indicators and sub-indicators are set, and then experts are invited to cooperate to build a comparison matrix and determine the weight distribution of indicators. Through data collection and index weight calculation, the evaluation index system is established and consistency test is carried out to ensure the reliability of the model. Finally, according to the scoring results of AHP, the rural tourism images can be optimized and the corresponding optimization scheme is put forward. For example, by analyzing the AHP scoring results, people can determine the key features and improvement direction of rural tourism images, thus optimizing the quality and attractiveness of the images. To sum up, it is the key step to realize the comprehensive evaluation and optimization of rural tourism images by using DenseNet model to perceive and extract features of rural tourism landscape images and grading the extracted features by AHP method. These steps not only provide a basis for the subsequent grading process, but also improve the accuracy and reliability of image evaluation, which provides important support for improving rural tourism experience and promoting sustainable development. The DenseNet model in this study is set in the rural tourist attractions of the Yellow River Basin in China. The model architecture is mainly based on DenseNet. By adding dense connections among each convolution layer, the feature learning ability of the model is effectively improved. In addition, the attention mechanism is introduced into the DenseNet model, which is mainly used to enhance the model’s ability to identify the key features of rural tourism landscape images. This mechanism is located at the output layer of each Dense Block and focuses on the important areas of the image by weighted average. The network structure of rural tourism image optimization based on DL proposed in this study is shown in Fig. 4. By combining convolution layer, pooling layer, attention mechanism and DenseNet’s dense connection layer, the key features of rural tourist attractions pictures are effectively extracted and classified through the full connection layer. With the introduction of attention mechanism, the network can focus more on the important part of the picture, which improves the model’s ability to identify and optimize the rural tourism image. The whole network structure is designed to improve the performance of the model while maintaining the effective use of computing resources. The network is mainly composed of the following parts: The input layer receives the preprocessed pictures of rural tourist attractions, which have a fixed size of 224 × 224 pixels and are normalized for network learning. The convolution layer uses a series of filters to extract the features in the picture, and each filter is responsible for capturing a specific feature, such as edges and textures. Pool layer is used to reduce the spatial size of feature map, reduce the amount of calculation, and keep important information at the same time. The commonly used pool operations are maximum pool and average pool. In order to make the model pay more attention to the key areas in the picture, attention mechanism is introduced into the network. This mechanism can dynamically adjust the attention of the model to different areas, so that the network can focus on areas with more information. The attention module is usually located behind the convolution layer and is realized by weighted summation of feature maps. The weights are learned from the model, which reflects the importance of each region. Dense connection layer is the core of DenseNet architecture, which allows each layer to be directly connected with all subsequent layers. This design improves the reuse rate of features, helps to alleviate the problem of gradient disappearance and improve network performance. In each densely connected layer, the output of the previous layer will be input to the beginning of all subsequent layers and merged with the outputs of all previous layers of these layers. The fully connected layer maps the learned advanced features to the final output space, such as the number of categories used for classification. The output layer uses Softmax activation function to classify multiple categories and output the probability distribution of each category. The loss function is used to measure the difference between the predicted value and the real value of the model, and the cross-entropy loss function is commonly used. The optimizer is responsible for updating the weights of the network to minimize the loss function. Adam optimizer is widely used because of its adaptive learning rate. Fig. 4Network structure of rural tourism image optimization based on DL. Network structure of rural tourism image optimization based on DL. Methodological approach This study employs DL models, particularly DenseNet, in combination with big data analysis and the AHP to explore the recognition and optimization of rural tourism images. The proposed methodology aims to integrate theoretical frameworks with real-world case studies by conducting an in-depth analysis of image data from specific rural tourism destinations, thereby achieving effective feature extraction and image evaluation. First, for image recognition, the study selects DenseNet as the foundational model, leveraging its densely connected structure to enhance feature extraction capabilities. When constructing the DenseNet model, the experiment finely tunes the network’s depth, growth rate, and compression factor to ensure the model can retain key information while avoiding overfitting when processing rural tourism images. Specifically, this study focuses on rural tourism destinations in the Yellow River Basin and collects a large volume of high-quality tourism images. These images undergo preprocessing, including resizing, normalization, and data augmentation, to ensure consistency and diversity in the input data. The processed images are then fed into the DenseNet model, where its densely connected architecture extracts features and passes information layer by layer, effectively overcoming the vanishing gradient problem and improving the model’s learning efficiency. Second, to better evaluate and optimize rural tourism images, the study adopts AHP to integrate various factors affecting image evaluation. By establishing a multi-level evaluation system that includes primary and sub-indicators, the experiment systematically identifies key factors influencing rural tourism image representation. The comparison matrix construction and weight distribution process, with expert involvement, ensures the scientific rigor and reliability of the model. Ultimately, based on AHP evaluation results, the study proposes targeted optimization strategies to enhance the appeal and expressiveness of rural tourism images. The research process also involves several other DL algorithms, such as ResNet, VGG, EfficientNet, and ViT, to conduct a comparative analysis of DenseNet’s performance. Each model has distinct advantages. ResNet addresses the challenges of deep network training with residual connections. VGG has strong feature extraction capabilities due to its simple structure. EfficientNet balances complexity and accuracy well. ViT handles high-resolution images robustly. The inclusion of these models not only enriches the depth and breadth of the study but also provides a comparison for DenseNet’s effectiveness, further validating the rationale behind its selection. Experimental design and performance evaluation Datasets collection In this study, test data of the rural tourism image perception model integrated with DenseNet are obtained from the official tourism website Mafengwo.com, with a total of more than 154,600 photos. The photos are in Joint Photographic Experts Group (JPEG) format with HD resolution: 1920*1080 pixels, 24-bit color depth, represented in the Red, Green, Blue (RGB) color space. The images in the dataset cover many types of landscapes, including mountain canyons, lakes and rivers, vegetation forests, farmland, village layout, orchards and gardens, and architectural relics. To address the issue of promotional images focusing predominantly on landscapes and facilities while overlooking aspects such as sanitary conditions, this study incorporates data from user feedback platforms, community reviews, and social media, particularly images related to the cleanliness and maintenance of tourist areas. These images primarily reflect the sanitary environment of the sites, waste disposal facilities, and the upkeep of public restrooms. This additional data is processed alongside the original landscape dataset and subjected to feature extraction and classification within the DenseNet model. During the image processing stage, specific annotations are added to images related to sanitary conditions and facility maintenance, ensuring that these data are fully utilized in model training. Moreover, on-site investigations are conducted to obtain further images and feedback on environmental maintenance, cleaning equipment, and waste classification. These measures significantly enhance the comprehensiveness of the dataset, providing a more accurate representation of the real conditions at rural tourist destinations. The number and distribution of images across different categories are shown in Table 1: Table 1Number and distribution of images in different categories.Image categoriesNumber of imagesAverage resolutionGeographic originMountains and valleys20,0001920*1080Yellow River coast, southern mountainous areasLakes and rivers15,0001920*1080Hunan, Hubei, Yellow River coastVegetation and forests25,0001920*1080Southwest China, Yellow River basinFarmland18,0001920*1080Rural areasVillage layout42,6001920*1080Rural areas in various provincesOrchards and gardens14,0001920*1080Orchards in the southArchitectural remains10,0001920*1080Historical sites in various placesHealth and environment related10,0001920*1080Collected from community feedback Table 1Number and distribution of images in different categories.Image categoriesNumber of imagesAverage resolutionGeographic originMountains and valleys20,0001920*1080Yellow River coast, southern mountainous areasLakes and rivers15,0001920*1080Hunan, Hubei, Yellow River coastVegetation and forests25,0001920*1080Southwest China, Yellow River basinFarmland18,0001920*1080Rural areasVillage layout42,6001920*1080Rural areas in various provincesOrchards and gardens14,0001920*1080Orchards in the southArchitectural remains10,0001920*1080Historical sites in various placesHealth and environment related10,0001920*1080Collected from community feedback Number and distribution of images in different categories. After testing the new dataset, the network model of rural tourism image optimization based on the labels in the dataset represent the scenic spots or landscape types corresponding to each photo. For example, the label “Mountain Canyon” means that the image is a photo of mountain canyon landscape, and “Village Layout” means that the image shows the spatial layout and landscape characteristics of the village. Each label has been manually annotated and explained to ensure the accurate description and classification of the image content. In addition, detailed information related to each label is provided, including the name, geographical location and feature description of the scenic spot. The process of relevant data is obtained from the DenseNet model. The image data obtained from Mafengwo.com is preprocessed, including the steps of image size adjustment and color space standardization. By using the model, feature extraction is carried out on photos of Mafengwo.com, and visual feature data for each photo is acquired. The DenseNet model is trained using a large Mafengwo.com photo dataset. A test image is identified by the DenseNet model as a specific scenic spot along the Yellow River, and the relevant data output by the model includes information such as the cognitive confidence score of the scenic spot. These data are employed to establish the basis for the rural tourism image evaluation and optimization model along the Yellow River. The relevant data on the rural tourism image along the Yellow River are obtained from the official website of the Bureau of Statistics, and the rural tourism image is evaluated and analyzed. The unique value of each village’s tourism image is then determined by combining these characteristic data with statistical information from the villages along the Yellow River. This includes information on the village’s visual attributes as well as other data points. Parameters setting The DL model used in this study, DenseNet-121, is trained and evaluated within a robust hardware and software environment. The experimental software environment includes the use of Place365-CNNs and SPSS version 26.0 for photo data classification and data analysis. The hardware setup consists of a Lenovo IdeaPad 15 (2022 version) with an Intel Core i9-9880XE CPU, featuring 18 cores and 36 threads, a base frequency of 3.0 GHz, and a boosted frequency of up to 4.4 GHz, along with 16 GB of RAM. No other dedicated computing devices are utilized aside from the CPU. A total of 150,000 images are processed in the experiment, and the VGG, ResNet, and DenseNet models are trained and evaluated to ensure that all experimental tasks are completed on this laptop. To enhance the training effectiveness of the model, the network parameters are optimized. The VGG model adopts VGG-16 architecture, which includes 16 convolution layers and 3 fully connected layers. Stochastic Gradient Descent (SGD) optimizer is used in training, and the initial learning rate is set to 0.001 and the momentum is 0.9. During the training process, if the accuracy of the verification set does not improve within 10 epoch, the learning rate will decay to 0.1 times of the previous learning rate. There are 50 epoch in the whole training process, and the batch size is 32. The selected ResNet model is the ResNet-50 version, which contains 50 layers of CNNs. SGD is also used as the optimizer, and the initial learning rate is 0.01 and the momentum is set to 0.9. The learning rate adjustment strategy is similar to VGG network. When the accuracy is stagnant, it will be attenuated. ResNet model trains a total of 70 epochs, and the batch size of each epoch is set to 64. The ratio of the training set to the test set is 6:4. In the training process, cross-validation is utilized to ensure the robustness of the model. The parameters of the network designed in this study: DenseNet-121 structure, including 121 layers. This study uses Adam optimizer to perform model training. The initial learning rate is set to 0.0001, and the default values of beta1 and beta2 (0.9 and 0.999 respectively) are used. With the increase of epoch, the learning rate gradually decays according to a predetermined strategy. For DenseNet, it sets up 100 epoch, and the batch size is selected as 64. 100 training iterations are conducted for each model to ensure that the model can fully learn the characteristics of the dataset and achieve convergence. In order to evaluate the stability and generalization ability of the model, five independent experiments on each DL model are conducted, that is, each model is independently trained and tested with different random initialization weights for five times, and finally the average of the five test results is taken to evaluate the model performance. The purpose of the experiment is to evaluate the performance of the proposed rural tourism image perception model based on DenseNet, and to establish a rural tourism image evaluation and optimization model based on big data analysis. The indicators such as accuracy, data loss rate and recall rate of different types of photos are used to measure the performance of the model. The meanings of these indicators are as follows: accuracy indicates the proportion of samples correctly classified by the model, data loss indicates the proportion of unsuccessful data extraction or retrieval, and recall indicates the ability of the model to correctly identify different types of photos. The recognition accuracy is depicted in Eq. (9):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Acc=\frac{{CCs}}{{{S_{total}}}} \times 100\%$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\text{C}\text{C}\text{s}$$\end{document} refers to the number of samples correctly classified. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{total}$$\end{document} indicates the total number of samples in the test set. The data loss rate reads:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Loss=\frac{{{S_{no}}}}{{{S_{total}}}} \times 100\%$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{S}}_{\text{n}\text{o}}$$\end{document} is the amount of data that has not been successfully extracted or retrieved for various reasons. This research work is a multi-classification problem, which needs to distinguish N categories of samples. In the multi-classification problem, the recall rate is calculated by weighted average. The weighted average takes into account the number of samples in each category, and gives greater weight to the categories with more samples to reflect the contribution of each category in the overall recall rate.11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${R_{weighted}}=\mathop \sum \limits_{{i=1}}^{N} \left( {\frac{{{S_i}}}{{{S_{total}}}} \times {R_i}} \right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{S}}_{\text{i}}$$\end{document} is the number of samples in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:i$$\end{document}-th category. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\text{S}}_{\text{t}\text{o}\text{t}\text{a}\text{l}}$$\end{document} is the total number of samples in the test set. Equation 12 shows the calculation of accuracy rate:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${P_{weighted}}=\mathop \sum \limits_{{i=1}}^{N} \left( {\frac{{{S_i}}}{{{S_{total}}}} \times {P_i}} \right)$$\end{document} The calculation of F1 score is shown in Eq. 13:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F{1_{weighted}}=2 \times \frac{{{P_{weighted}} \times {R_{weighted}}}}{{{P_{weighted}}+{R_{weighted}}}}$$\end{document} The correlation between the image of rural tourism and the location is calculated in Eq. (14):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$C=\frac{{Covariance\left( {IM,LO} \right)}}{{SD\left( {IM} \right)*SD\left( {LO} \right)}} \times 100\%$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Covariance(IM,LO)$$\end{document} represents the covariance between rural tourism image and location. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:SD\left(IM\right)$$\end{document} is the standard deviation (SD) of rural tourism image, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:SD\left(LO\right)$$\end{document} means the SD of rural tourism location. The established rural tourism image evaluation system realizes the rural tourism image evaluation based on big data, as exhibited in Table 2: Table 2The rural tourism image evaluation system.System indexesRural tourism imagePrimary indexDegree of authenticity ASecondary indexHistorical sitesSpatial patternTraditional festivalsCultural experienceExpressionA1A2A3A4Primary indexAesthetics BSecondary indexComprehensive harmonyVisual pollutionLandscape environmentColor qualityExpressionB1B2B3B4Primary indexService facility CSecondary indexAccommodation facilitiesCatering servicesTransportation facilitiesFunctional facilitiesExpressionC1C2C3C4Primary indexManagement situation DSecondary indexPublic receptionEmergency managementCultural atmosphereTargeted servicesExpressionD1D2D3D4Primary indexSafety level ESecondary indexSocial securityFacility safetyProtective measuresTraffic capacityExpressionE1E2E3E4Primary indexImage communication FSecondary indexDiversified channelsDiverse contentCommunication timelinessAuthentic contentExpressionF1F2F3F4 Table 2The rural tourism image evaluation system.System indexesRural tourism imagePrimary indexDegree of authenticity ASecondary indexHistorical sitesSpatial patternTraditional festivalsCultural experienceExpressionA1A2A3A4Primary indexAesthetics BSecondary indexComprehensive harmonyVisual pollutionLandscape environmentColor qualityExpressionB1B2B3B4Primary indexService facility CSecondary indexAccommodation facilitiesCatering servicesTransportation facilitiesFunctional facilitiesExpressionC1C2C3C4Primary indexManagement situation DSecondary indexPublic receptionEmergency managementCultural atmosphereTargeted servicesExpressionD1D2D3D4Primary indexSafety level ESecondary indexSocial securityFacility safetyProtective measuresTraffic capacityExpressionE1E2E3E4Primary indexImage communication FSecondary indexDiversified channelsDiverse contentCommunication timelinessAuthentic contentExpressionF1F2F3F4 The rural tourism image evaluation system. In Table 2, the established rural tourism image evaluation system has a total of 6 primary indexes and 24 secondary indexes. Based on this, the image of rural tourism is analyzed, and the direction of rural tourism image optimization is proposed according to specific examples. Performance evaluation Figure 5 suggests the change in data loss and accuracy of the training set as the number of training times increases during model training: Fig. 5Training of DenseNet model. Training of DenseNet model. Figure 5 details that in the process of model training, with the increase of training times, the accuracy of the DenseNet model continues to improve, and finally stabilizes above 92% and reaches 94.73%, illustrating the model’s accuracy. The data loss rate is less than 0.02%, which ensures the integrity of the data to the greatest extent, and guarantees the data characteristics and high trustworthiness of the rural tourism image analysis. The correlation between rural tourism image perception and location space is outlined in Fig. 6: Fig. 6The correlation degree between rural tourism image perception and location space ((a) represents the relationship between quality and location; (b) denotes the relationship between uniqueness and location). The correlation degree between rural tourism image perception and location space ((a) represents the relationship between quality and location; (b) denotes the relationship between uniqueness and location). Figure 6 illustrates that this study adopts correlation analysis and compares the global correlation coefficient according to the perception mechanism of rural tourism image to judge the degree of correlation. Through comparison, it is found that the correlation coefficient between visual aesthetic quality and location is -0.11473, indicating that the two have a negative correlation. The results indicate that location is not a decisive factor in aesthetic perception. The possible reason lies in the profound impact that the cultural atmosphere, landscape design, and environmental protection measures of different regions have on visitors’ visual experiences. Future research could introduce more dimensional variables, such as community involvement, marketing strategies, and environmental sustainability, to better understand how these factors interact in various tourism contexts, thereby influencing tourists’ overall perception of rural tourism images. The correlation coefficient between visual uniqueness and location is 0.185, which is still relatively weak. This suggests that visual uniqueness is not the only factor affecting the image of rural tourism. Other factors, such as accessibility, local service quality, cultural and historical context, and protective measures, may have a more significant influence on visitors’ image perception. Especially in some rural areas, the convenience of access for tourists and the completeness of service facilities directly impact their overall evaluation of rural tourism images. Therefore, future research will explore how these external factors interact with visual characteristics to provide a comprehensive understanding of the construction of rural tourism images. For instance, improvements in transportation infrastructure may enhance the impact of visual uniqueness on visitor attraction, while optimizing service facilities can elevate the perception of visual aesthetics. This shows that the relationship between the two is a positive correlation. Through comparative analysis, it is found that establishing a tourism image needs to consider the visual uniqueness of local rural tourism. Figure 7 portrays the change in recall of different types of photos in the training and test sets. Among them, mountain canyons, lakes and rivers, vegetation and forests, farmland, village patterns, orchards and gardens, and architectural relics are recorded as seven indicator types H1-H7, and the total recall is represented by H. Fig. 7The variation of recall of diverse types of photos. The variation of recall of diverse types of photos. In Fig. 7, recall is generally lower than the test data during training. In all data testing processes, recall can generally reach over 74%, and the model has a high recognition ability for real examples. Figure 8 shows the consistency test and evaluation weight of various indicators of rural tourism image. The weights and consistency test of 6 primary indexes and 24 secondary indexes are selected to evaluate the rural tourism image, and the number of evaluators is determined. Fig. 8The weight and consistency test of rural tourism image evaluation indicators ((a) indicates the weight and consistency test of the primary index; (b) Indicates the weight of the secondary index and the number of people evaluated.) The weight and consistency test of rural tourism image evaluation indicators ((a) indicates the weight and consistency test of the primary index; (b) Indicates the weight of the secondary index and the number of people evaluated.) Figure 8 describes that the degree of authenticity and service facilities have the greatest effect on the rural tourism image, and the weight scores are 0.21555 and 0.25882, which indicates that the characteristics of local rural tourism and tourism services are extremely important for the establishment of tourism image. The consistency test demonstrates that the random consistency rate of all indicators is less than 0.1, which meets the credibility requirements and is recorded as 1, so rural tourism image evaluation could be carried out. Figure 9 depicts the evaluation results of the rural tourism image of six representative locations in the selected Yellow River basin, namely Shenquan Village (L), Zhangjiazhuang (M), Yukou Village (N), Nihegou (P), Muyu (Q), and Heyeping (R). Fig. 9Evaluation results of tourism image of representative villages in the Yellow River basin. Evaluation results of tourism image of representative villages in the Yellow River basin. Figure 9 indicates that from the perspective of vertical comparison, Nihegou (P) has a high degree of authenticity and service facilities, with scores of 3.7275 and 4.2168. Combined with the weight representation, the rural tourism image of this area is the most distinctive. Previous studies have explored similar datasets and used different DL models for experiments. For example, Cho et al. (2022) used Residual Network (ResNet) model to analyze and optimize similar rural tourism landscape images57. In addition, Wang et al. (2023) tried the application of VGG model on the same data set and achieved certain results. By summarizing and comparing these studies, people can evaluate the effectiveness and applicability of the method proposed in this study more comprehensively58. In order to further verify the validity and applicability of the rural tourism landscape image optimization method proposed in this study two advanced DL models, EfficientNet and Vision Transformer (ViT), are added as additional comparison objects in the experiment. EfficientNet model has attracted wide attention because of its efficient scaling strategy and excellent performance, while ViT model has attracted much attention because of its outstanding performance on large-scale image datasets. To comprehensively evaluate the performance of the proposed DenseNet-based model for rural tourism image perception, this study further introduces additional evaluation metrics, including Top-K accuracy, confusion matrix, Cohen’s Kappa, No Information Rate, sensitivity, and specificity, building upon the original accuracy, data loss rate, and recall rate. After experimental comparison on the same dataset, the results show that EfficientNet and ViT are superior to ResNet and VGG models in image recognition accuracy, and have similar performance indexes compared with DenseNet. The results are shown in Table 3 below. The DenseNet model achieves the highest accuracy of 94.73% with the lowest data loss rate of only 0.02%, demonstrating its efficiency and stability in data processing. The recall rate is 74.5%, the precision reaches 95.20%, and the F1 score is 84.60%, further confirming its excellent performance in classification tasks. In comparison, the accuracy of the ResNet and VGG models is 93.21% and 91.85%, respectively, with slightly higher data loss rates of 0.03% and 0.04%. The EfficientNet and ViT models closely follow in terms of accuracy, recording 94.12% and 94.57%, respectively, and they also exhibit excellent data loss rates, particularly EfficientNet at just 0.01%. Additionally, the ViT model excels in recall (78.1%), precision (95.80%), and F1 score (87.00%), approaching the performance level of DenseNet. In terms of Cohen’s Kappa, ViT achieves a value of 0.90, slightly higher than DenseNet’s 0.89, indicating a greater consistency between its classification results and the actual labels. Overall, DenseNet demonstrates balanced performance across accuracy, data loss rate, and comprehensive evaluation metrics, showcasing its efficiency and reliability in optimizing rural tourism images. Table 3Comparison results of different DL algorithms on the same dataset.ModelAccuracy (%)Data lossRecall rate (%)Precision (%)F1 value (%)Top-3 accuracy (%)Cohen’s KappaNIR(%)Sensitivity (%)Specificity (%)DenseNet94.730.0274.595.2084.6098.000.89859093ResNet93.210.0372.894.1082.8096.500.85858891VGG91.850.0470.292.8080.5095.200.80858589EfficientNet94.120.0176.394.5085.2097.800.87858992ViT94.570.0278.195.8087.0098.500.90859194With the rise of rural tourism, promoting its image optimization has become an important issue. This study verifies the effectiveness of the network model of rural tourism image optimization based on DL in this application scene by introducing a variety of image datasets of rural tourism areas. It includes mountains, hills, plains and other landscapes with different landforms, as well as rural scenery, ancient villages, folk activities and other elements. The dataset covers representative rural tourist destinations such as fishing villages in the southeast coast of China, rural areas in the northwest plateau, tea gardens in the middle and water towns in the south. The testing results of different types of rural tourism scenic image optimization network models are shown in Table 4: Table 3Comparison results of different DL algorithms on the same dataset.ModelAccuracy (%)Data lossRecall rate (%)Precision (%)F1 value (%)Top-3 accuracy (%)Cohen’s KappaNIR(%)Sensitivity (%)Specificity (%)DenseNet94.730.0274.595.2084.6098.000.89859093ResNet93.210.0372.894.1082.8096.500.85858891VGG91.850.0470.292.8080.5095.200.80858589EfficientNet94.120.0176.394.5085.2097.800.87858992ViT94.570.0278.195.8087.0098.500.90859194With the rise of rural tourism, promoting its image optimization has become an important issue. This study verifies the effectiveness of the network model of rural tourism image optimization based on DL in this application scene by introducing a variety of image datasets of rural tourism areas. It includes mountains, hills, plains and other landscapes with different landforms, as well as rural scenery, ancient villages, folk activities and other elements. The dataset covers representative rural tourist destinations such as fishing villages in the southeast coast of China, rural areas in the northwest plateau, tea gardens in the middle and water towns in the south. The testing results of different types of rural tourism scenic image optimization network models are shown in Table 4: Comparison results of different DL algorithms on the same dataset. With the rise of rural tourism, promoting its image optimization has become an important issue. This study verifies the effectiveness of the network model of rural tourism image optimization based on DL in this application scene by introducing a variety of image datasets of rural tourism areas. It includes mountains, hills, plains and other landscapes with different landforms, as well as rural scenery, ancient villages, folk activities and other elements. The dataset covers representative rural tourist destinations such as fishing villages in the southeast coast of China, rural areas in the northwest plateau, tea gardens in the middle and water towns in the south. The testing results of different types of rural tourism scenic image optimization network models are shown in Table 4: Table 4Test results of image optimization network models for different types of rural tourist attractions.Rural landscape typeDataset sizeAccuracy (%)Data loss rate (%)Callback rate (%)Top-3 accuracy (%)Sensitivity (%)Specificity (%)Mountains100092590968891Hills80089688938589Plain1200934.591958990Rural scenery1500943.892979093Ancient village700905.789948688Folk activities600915.290958789Sanitary facilities500886.587928486Facility maintenance500896.288938587 Table 4Test results of image optimization network models for different types of rural tourist attractions.Rural landscape typeDataset sizeAccuracy (%)Data loss rate (%)Callback rate (%)Top-3 accuracy (%)Sensitivity (%)Specificity (%)Mountains100092590968891Hills80089688938589Plain1200934.591958990Rural scenery1500943.892979093Ancient village700905.789948688Folk activities600915.290958789Sanitary facilities500886.587928486Facility maintenance500896.288938587 Test results of image optimization network models for different types of rural tourist attractions. After testing the new dataset, the network model of rural tourism image optimization based on DL shows its powerful feature learning ability. Further, the research team selects some rural tourism destinations for on-the-spot investigation, and collects the corresponding feedback pictures from tourists. The model demonstrates high accuracy across various landscape types, particularly excelling in the categories of rural scenery (94%) and plains (93%), with data loss rates of 3.8% and 4.5%, respectively, and recall rates reaching 92% and 91%. The accuracy rates for the mountain and hill categories are 92% and 89%, respectively; although their data loss rates are slightly higher, they remain within acceptable limits. The accuracy for the ancient villages and folk activities categories is 90% and 91%, respectively, indicating that the model also possesses good recognition capabilities when dealing with culturally distinctive landscapes. However, the accuracy rates for the sanitation facilities and facility maintenance categories are relatively lower, at 88% and 89%, with data loss rates of 6.5% and 6.2%, respectively, suggesting that the model’s recognition performance in these areas needs further improvement. Overall, the model performs excellently across most rural landscape types, effectively supporting the optimization of rural tourism images. However, it still needs to strengthen its capabilities in specific areas such as environmental sanitation and facility maintenance to enhance overall recognition accuracy and the model’s practicality. In summary, the DenseNet model demonstrates significant advantages in optimizing rural tourism images, which holds important practical implications for the development of rural tourism. The high accuracy and low data loss rates ensure the reliability of image classification, aiding in the accurate representation of the diversity and uniqueness of rural tourism resources. The improvement in recall rates signifies that the model has a higher coverage rate when recognizing different types of rural landscapes, ensuring comprehensive image analysis. In terms of rural tourism policy formulation, the findings can provide scientific decision-making bases for relevant departments. Through precise image classification and analysis, policymakers can gain a better understanding of the distribution and characteristics of various landscapes, allowing for targeted tourism promotion strategies. For instance, areas identified as having high visual appeal and uniqueness can be prioritized for publicity and infrastructure development to enhance overall tourism attractiveness. Moreover, the model’s ability to recognize environmental sanitation and facility maintenance issues aids in promptly identifying and addressing related problems, ensuring a positive experience for visitors and promoting the sustainable development of rural tourism. Additionally, the confusion matrix and Cohen’s Kappa value further validate the classification consistency of the DenseNet model across different landscape types. The DenseNet model exhibits high classification accuracy in most categories, particularly standing out in the “gorge” and “lakes and rivers” categories, showcasing its advantages in processing natural landscape images. Although some confusion exists in the “farmland” and “architectural relics” categories, the overall classification performance remains satisfactory, indicating that the DenseNet model possesses strong adaptability and robustness. The DenseNet model not only outperforms other models in technical metrics but also demonstrates significant potential in practical applications. The results of this study can provide effective technical support for rural tourism image optimization, promoting scientific management and sustainable development of rural tourism. The influence of rural image promotion on sustainable development This study is devoted to using image recognition technology to enhance rural image and promote the sustainable development of rural tourism. The application of image recognition technology is very important for improving the efficiency of rural scenic spot management and optimizing the tourist experience, thus attracting more tourists, increasing local income and promoting the vigorous development of rural economy. Specifically, by identifying and classifying the images of rural scenic spots, people can find and solve the problems existing in the management of scenic spots in time, such as environmental sanitation and facility maintenance to enhance the overall image and quality of scenic spots. In addition, through the analysis of the photos taken by tourists in scenic spots, people can deeply understand the preferences and needs of tourists, improve the facilities and services in scenic spots and enhance the satisfaction of tourists. This further enhances the return rate and word-of-mouth effect of tourists and lays a solid foundation for the sustainable development of scenic spots. In the application of image recognition technology, people can not only monitor and manage the environment and facilities of scenic spots, but also finely optimize and adjust the services and facilities of scenic spots according to the actual behaviors and experiences of tourists. Through the intelligent analysis of the photos taken by tourists in the scenic spot, people can understand the tourists’ evaluation and feelings about different aspects of the scenic spot to improve and upgrade the services and facilities of the scenic spot in a targeted manner and meet the diversified needs of tourists. This personalized service and continuous improvement can not only enhance the overall experience of tourists, but also increase tourists’ praise and recommendation for scenic spots, thus driving more tourists to come and realizing the sustainable development goal of rural tourism. To sum up, the image recognition technology in this study is of great significance in enhancing the image of rural areas and has played a positive role in promoting the sustainable development of rural tourism. By optimizing the management and service of rural scenic spots, not only can the overall image and quality of scenic spots be improved, but also the satisfaction and loyalty of tourists can be enhanced, and the virtuous cycle development of rural tourism can be realized. Disscussion Researchers both domestically and internationally have conducted extensive studies on optimizing the image of rural tourism. They actively explore innovations in areas such as smart tourism experiences, landscape recognition, social media analysis, and sustainable development, providing a wealth of ideas and methods for the advancement of rural tourism. However, previous research has primarily focused on different aspects of optimizing rural tourism images, such as environmental monitoring, image recognition, and promotional strategies, lacking a comprehensive framework that integrates IoT and DL technologies. Moreover, the influence of big data on decision-making and assessment has not been fully considered. Therefore, this study aims to establish an integrated model that combines IoT and DL technologies to optimize rural tourism images, ultimately enhancing visitor experiences and promoting sustainable development in the industry. This study optimizes the image of rural tourism by combining IoT and DL technologies to enhance visitor experience and drive sustainable development. Zhang developed an intelligent tour guide system based on IoT and DL technologies to provide tourists with real-time scenic spot information, historical and cultural background, and personalized recommendations through sensors and mobile applications59. Yang (2021) explored combining IoT technology with VR and AR to create a richer travel experience for tourists, comprising virtual historical scenes, cultural recreation, and other content60. Tan (2021) focused on analyzing tourists’ comments and feedback on social media through DL technology, understanding tourists’ emotional tendencies and opinions, and offering useful information to optimize and improve rural tourism image61. Nguyen-Da et al. (2023) proposed a hybrid algorithm combining CNN and Long Short-Term Memory (LSTM) to accurately predict the tourism demand in Vietnam and some provinces62. The results revealed that CNN-LSTM achieved excellent performance on various evaluation benchmarks, which indicated that the combination of CNN and LSTM algorithms had better effectiveness and higher reliability. This study used the same DenseNet-implemented model, and impressive outcomes were obtained in the area of picture recognition using DL technology. Su et al. divided the image types of the image data in the comments by using the image semantic segmentation method based on deep CNNs and calculated the proportion of the types63. The results showed that according to the proportion calculated by word frequency, the learning value accounted for 55.26%, the cultural and historical value accounted for 30.92%, the entertainment value, the aesthetic value, and the health value accounted for 8.88%, 3.29%, and 1.65%, respectively. This image analysis method contributed to understanding the visual characteristics of rural tourism images. Unlike the model constructed by DenseNet used here, Su et al. mainly focused on semantic segmentation and type division of images, while this study paid more attention to overall image evaluation and optimization based on big data. Bravo et al. used linear regression, CNN regression, decision trees, and random forests to predict the number of domestic and foreign tourists64. The results indicated that linear regression could better predict the number of tourists and realize tourism’s digital transformation. Though the emphasis was different, this was connected to the big data analysis and evaluation paradigm used in this study. This study concentrated on improving the perception of rural tourism, whereas Bravo et al. were more concerned in forecasting visitor numbers to direct the digital transition. In the end, Alsayat (2023) developed hybrid approaches using K-means, latent Dirichlet al.location (LDA), sequential minimum optimization (SMO), and support vector regression. The satisfaction of each segment was revealed based on travelers’ online comments about the hotel65. The results illustrated that the method was effective for the big data analysis and traveler segmentation of Mecca hotels. Although the tourism background and specific methods of this study were different, they all involved the application of DL and big data methods in tourism research. Together, these studies highlighted the importance of comprehensive research in the field of tourism, combining advanced technologies and DL methods, to provide strong support for improving the visitor experience and promoting the sustainable development of tourism. The analysis of the rural tourism image perception model indicates that the accuracy of the DenseNet model improves with an increasing number of training iterations, stabilizing above 92% and reaching a peak of 94.73%. This research employs large-scale data analysis methods and applies the DenseNet algorithm to rural tourism image perception for the first time, achieving a comprehensive evaluation and optimization of the images. These findings not only provide practical guidance for the rural tourism industry but also offer new ideas and methods for the technological development and application in related fields. Conclusion This study aims to optimize rural tourism images, enhance visitor experiences, and promote sustainable development by integrating IoT and DL technologies. The results indicate that the DenseNet model demonstrates significant advantages in optimizing rural tourism images, achieving an accuracy of 94.73% and a data loss rate of only 0.02%, showcasing its efficiency and stability in processing complex image data. In comparison, other DL models such as ResNet, VGG, EfficientNet, and ViT also exhibit good performance. However, the balance of DenseNet across comprehensive metrics makes it a superior choice. This result is primarily attributed to DenseNet’s densely connected architecture, which effectively extracts image features and mitigates the gradient vanishing problem, thus achieving higher accuracy and robustness in rural tourism image classification tasks. Additionally, the weak negative correlation (−0.11473) observed between visual aesthetic quality and location, along with the weak positive correlation (0.185) between visual uniqueness and location, suggests that location is not the sole determining factor of rural tourism images. Instead, multiple factors such as cultural atmosphere, landscape design, environmental protection measures, accessibility, local service quality, and marketing strategies collectively influence visitors’ perceptions of rural tourism images. This finding emphasizes the need to consider various factors comprehensively when formulating rural tourism development policies to achieve rational development and protection of tourism resources, thereby enhancing overall tourism appeal. The application of IoT technology in this study provides a solid foundation for collecting and analyzing image data through real-time monitoring of environmental factors and visitor flow. Coupled with the automated recognition and analysis capabilities of DL models, this not only improves the efficiency and accuracy of image classification but also supports scientific management of rural tourism with data insights. This technological integration demonstrates its immense potential in promoting sustainable tourism development, enabling optimal resource allocation and effective environmental protection. However, this study also faces challenges, such as the need to further enhance data diversity and the model’s generalization ability. Future research could introduce more dimensional variables to explore the interaction mechanisms between different factors, further refining the theoretical framework for optimizing rural tourism images. Furthermore, as IoT and DL technologies continue to advance, their application prospects in rural tourism will broaden, providing new pathways and methods for achieving smart tourism. In summary, this study not only enriches the theoretical framework for rural tourism optimization but also provides practical technical support for relevant policy formulation and implementation, advancing the scientific management and sustainable development of rural tourism.