The optimization path of agricultural industry structure and intelligent transformation by deep learning Authors: Xingchen Pan, Jinyu Chen Published: 2024 ABSTRACT This study addresses key challenges in optimizing agricultural industry structures and facilitating intelligent transformation through the application of deep learning algorithms and advanced optimization techniques. An intelligent system for agricultural industry optimization is developed, with convolutional neural networks, recurrent neural networks, Long Short-Term Memory networks, and generative adversarial networks introduced for tasks such as image recognition, time series forecasting, and synthetic data generation. Subsequently, a hybrid optimization method is designed, combining the Genetic Algorithms with particle swarm optimization to improve the model’s global search capability and local convergence speed. The performance of these techniques is rigorously evaluated through extensive experimentation. The results demonstrate that the proposed method outperforms conventional algorithms in regression tasks, particularly in terms of computational efficiency, data processing speed, and model training stability, while also exhibiting high scalability. In crop yield prediction, the proposed method achieves superior performance, as evidenced by reductions in both absolute error and mean squared error, along with attaining the highest R2 value (0.93). Additionally, in pest and disease detection, the proposed method exceeds other models in accuracy (97.5%), precision (96.8%), recall (97.2%), and F1 score (0.97), underscoring its superior performance in detecting agricultural pests and diseases. The method also significantly surpasses traditional algorithms in crop disease identification accuracy, climate change prediction precision, and the quality of synthetic data generation. This study offers novel technical solutions and decision-making tools for advancing intelligent agriculture. Introduction Research background and motivations Despite some progress driven by technological innovation, the global agricultural industry continues to face several challenges. Firstly, agricultural data is highly complex and diverse, encompassing multidimensional information such as climate, soil, and crop growth. Traditional analytical methods struggle to manage the nonlinear relationships between these complex datasets. Secondly, while existing deep learning models perform well in specific scenarios, their generalization ability is often limited, particularly in agricultural applications that span different regions and crop types. Finally, the dynamic nature of agricultural production requires systems that can adjust models and provide real-time feedback. However, most current methods lack this crucial mechanism, leading to delays and a lack of adaptability in decision-making1,2. Although some studies focus on waste resource management and pollution control, there is a significant gap in research on agricultural industry structure optimization, particularly in the integration of advanced deep learning technologies and intelligent management methods. Currently, deep learning technology, with its powerful data processing and pattern recognition capabilities, is increasingly being applied to the agricultural sector, facilitating the transition from traditional agriculture to smart agriculture, optimizing resource utilization, and reducing environmental impacts3,4. However, existing research on the application of deep learning in agricultural industry structure optimization and intelligent transformation is still incomplete, lacking comprehensive solutions that address resource management, sustainable development, and production efficiency5,6. Therefore, an innovative, deep learning-based systematic solution is urgently needed to address the issues of resource waste and environmental pollution in agriculture, which is the primary goal of this study. Research objectives The primary objective of this study is to propose a deep learning-based approach for optimizing the agricultural industry structure, with the aim of addressing resource waste and environmental pollution in the agricultural sector. Specifically, the objectives are as follows: Develop deep learning models to handle multidimensional data in agricultural production, improving the accuracy of crop yield predictions.Integrate intelligent management methods to optimize resource utilization in agricultural production and reduce environmental impact. Develop deep learning models to handle multidimensional data in agricultural production, improving the accuracy of crop yield predictions. Develop deep learning models to handle multidimensional data in agricultural production, improving the accuracy of crop yield predictions. Integrate intelligent management methods to optimize resource utilization in agricultural production and reduce environmental impact. Integrate intelligent management methods to optimize resource utilization in agricultural production and reduce environmental impact. A hybrid optimization algorithm is designed and evaluated to enhance the model’s global search capability and local convergence ability, verifying its effectiveness in promoting agricultural sustainability. Research plan and focus The main innovations of this study are reflected in the following three aspects: Application of Hybrid Optimization Algorithms: Traditional deep learning models often face challenges such as local optima and slow convergence in parameter adjustment and resource optimization. This study combines genetic algorithms (GAs) and particle swarm optimization to form a hybrid optimization algorithm, which effectively integrates global search and local optimization during model training, improving resource utilization efficiency and reducing computational costs. This innovation addresses the issue of local optima in processing agricultural data. Integration of Adaptive Learning and Real-Time Feedback Mechanisms: Agricultural production is influenced by dynamic factors such as climate and soil conditions, requiring real-time model parameter adjustments. This study introduces an adaptive learning algorithm and a real-time feedback mechanism, enabling the model to dynamically adjust parameters in response to changes in the external environment and optimize decision-making processes in real-time. This innovation effectively resolves the current limitations of agricultural models in quickly responding to environmental changes. Cross-Domain Data Fusion and the Application of generative adversarial networks (GANs): Agricultural data is highly diverse and unevenly distributed, and existing methods often rely on specific datasets, leading to insufficient generalization. This study introduces GANs to generate synthetic agricultural data, enhancing the diversity of training data and improving the generalization capability of the model. It addresses the problem of data scarcity and poor generalization, and validates the effectiveness of its application across different agricultural scenarios. The structure of this study is as follows: “Literature Review” Section reviews the relevant literature and analyzes the shortcomings of existing research. “Research Model” Section describes in detail the proposed research methods, including the design of deep learning models, hybrid optimization algorithms, and real-time feedback mechanisms. “Experimental Design and Performance Evaluation” Section presents the experimental design, data collection, and model performance evaluation, along with a discussion of the results and their implications. Finally, “Conclusion” Section provides conclusions and suggestions for future research directions. Literature review In recent years, deep learning technologies have demonstrated significant potential within the agricultural sector7. These advancements have contributed to major progress in areas such as pest and disease detection, crop growth forecasting, and soil quality evaluation8. For instance, Veerachamy and Ramar (2022) utilized convolutional neural networks (CNNs) for image recognition, enabling farmers to accurately diagnose crop diseases and implement timely preventive actions9. Similarly, Attri et al. (2023) highlighted the effectiveness of recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks in agricultural time series analysis, which are crucial for predicting the effects of climate change on crop yields, thus offering critical insights for agricultural decision-making10. With the successful application of metaheuristic algorithms in optimization models, the performance of deep learning has seen further enhancement. For instance, Jovanovic et al. (2024) proposed a framework for crop yield prediction based on metaheuristic algorithm-tuned agnostic neural networks, significantly improving the model’s prediction accuracy11. Additionally, Kolipaka and Namburu (2024) designed an automated crop yield prediction framework using a two-stage classifier, optimizing its performance through metaheuristic methods12. These successful cases demonstrate that metaheuristic algorithms can effectively enhance the performance of deep learning models in agriculture, addressing the common issue of traditional models being prone to local optima. Although deep learning has achieved significant progress in agricultural applications, research on optimizing agricultural industry structures remains relatively limited. Most existing studies focus on promoting sustainable development by improving resource allocation and enhancing agricultural efficiency13,14. The continuous advancement of modern agricultural technologies has facilitated the optimization and upgrading of agricultural structures, including improvements in planting configurations, processing of agricultural products, and distribution systems15. Sharma et al. (2022) underscored the importance of integrating regional agricultural characteristics with sustainable development strategies, proposing various optimization models to address agricultural production challenges in different regions16. For instance, Corceiro et al. (2023) employed a graph CNN optimized with metaheuristic algorithms to classify carrot grading systems, significantly enhancing the accuracy of agricultural product quality assessment17. Similarly, Kathole et al. (2024) advanced the intelligent optimization of the agricultural industry through the identification and classification of pests and diseases using a metaheuristic deep learning approach18. As a crucial component of agricultural modernization, smart agriculture integrates the Internet of Things (IoT), big data, and artificial intelligence (AI) technologies to enable intelligent and precise management of agricultural production19. The extensive use of intelligent sensor networks provides real-time monitoring and precision control capabilities, which in turn improve resource utilization efficiency and agricultural product quality20. Additionally, smart agriculture accelerates the digitization of agriculture, facilitating the optimization and intelligent upgrading of the agricultural supply chain21. Seyedmohammadi et al. (2023) proposed a crop yield prediction model based on an echo state network optimized by metaheuristic algorithms, significantly enhancing the model’s prediction accuracy and scalability through metaheuristic optimization of the network structure22. In recent years, the application of deep learning in agriculture has increasingly focused on multimodal fusion and small-object detection in complex environments to meet the demands of real-world scenarios. For instance, You and Liu (2023) optimized small-object detection using an improved You Only Look Once version 5 (YOLOv5) algorithm. By incorporating a Context-Aware Spatial Module (CASM) attention mechanism and multi-scale feature learning, they significantly improved the model’s accuracy in detecting small objects within complex backgrounds while reducing computational complexity, which provided new insights for small-object detection in agriculture23. Additionally, You et al. (2024) proposed a multimodal fusion approach that integrated physical information, such as sound and vibration, into deep learning models to enhance interpretability and stability in diagnosing equipment faults in agricultural machinery. This method, based on constraint-guided multimodal information fusion, enabled the model to more accurately capture environmental features, leading to improved diagnostic accuracy and reduced computational demands. This approach showed considerable potential for environmental monitoring in agricultural applications24. These studies provide effective solutions for complex environment perception and multimodal data fusion in smart agriculture and serve as valuable references for combining deep learning with multimodal optimization in this study. Despite the significant advancements in smart agriculture and deep learning technologies, several limitations remain. First, many studies are overly reliant on specific datasets and algorithms, resulting in models with limited generalization capabilities. Second, although some research employs metaheuristic algorithms to optimize models, these applications are often restricted to particular crop types or geographic regions, lacking broad validation across large-scale, cross-regional agricultural data. Moreover, existing studies on metaheuristic methods tend to focus on optimizing individual tasks or specific parameters, without offering a systematic solution for the overall optimization of agricultural industry structures. In response to these challenges, this study proposes a comprehensive optimization pathway for agricultural industry structures based on deep learning. By integrating advanced deep learning techniques with intelligent management methods, the goal is to improve agricultural production efficiency, minimize resource waste, and promote environmentally sustainable development. The innovation of this study lies in the combination of deep learning with hybrid optimization techniques, offering a systematic solution that is validated through practical applications to demonstrate its effectiveness in enhancing agricultural sustainability. Research model This study introduces an intelligent agricultural industrial structure optimization system that leverages deep learning and optimization technologies. The system aims to mitigate issues of resource waste and environmental pollution within the agricultural sector while enhancing production efficiency and sustainability. Key methodologies employed include CNNs, RNNs, their variant LSTM networks, GANs, and hybrid optimization techniques such as GAs and particle swarm optimization. The flow chart of this method is shown in Fig. 1: Fig. 1The process of the proposed method. The process of the proposed method. Application of deep learning algorithms in agriculture CNNs are primarily utilized for tasks such as crop disease and pest recognition, land cover classification, and various other image recognition applications25,26. The selection of CNNs is predicated on their exceptional performance in image processing, which facilitates the automatic extraction and classification of image features. The fundamental architecture of CNNs comprises convolutional layers, pooling layers, and fully connected layers. The convolution operation is mathematically represented by 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}$$\:\left(f\text{*}g\right)\left(t\right)={\int\:}_{-{\infty\:}}^{{\infty\:}}f\left(\tau\:\right)g(t-\tau\:)d\tau\:$$\end{document} In Eq. (1), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t$$\end{document} represents the time step, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f$$\end{document} represents the function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g$$\end{document} is the convolution kernel, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document} is the integration variable. Multiple CLs are used to extract image features, and the FCLs are used to achieve classification. RNNs and LSTM networks are predominantly employed for predicting time series data, including climate change patterns and crop growth cycles27,28. While RNNs are capable of capturing temporal dependencies within sequential data, they face challenges such as the vanishing gradient problem. LSTM effectively mitigates this issue by integrating input gates, forget gates, and output gates into their architecture. The fundamental unit formula for RNNs is represented by 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}$$\:{h}_{t}=\text{t}\text{a}\text{n}\text{h}({W}_{xh}{x}_{t}+{W}_{hh}{h}_{t-1}+{b}_{h})$$\end{document} In Eq. (2), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t}$$\end{document} represents the hidden state; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{xh}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{hh}$$\end{document} are the weight matrices; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{h}$$\end{document} denotes the bias. The LSTM architecture features input gates, forget gates, and output gates, which are specifically designed to address the vanishing gradient problem frequently encountered in RNNs29,30. Figure 2 illustrates the structure of an LSTM network: Fig. 2LSTM structure. LSTM structure. The update of LSTM is shown in Eqs. (3)-(8):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{t}=\sigma\:({W}_{f}\cdot\:[{h}_{t-1},{x}_{t}]+{b}_{f})$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{i}_{t}=\sigma\:({W}_{i}\cdot\:[{h}_{t-1},{x}_{t}]+{b}_{i})$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{C}}_{t}=\text{t}\text{a}\text{n}\text{h}({W}_{C}\cdot\:[{h}_{t-1},{x}_{t}]+{b}_{C})$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{t}={f}_{t}\text{*}{C}_{t-1}+{i}_{t}\text{*}\stackrel{\sim}{{C}_{t}}$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{o}_{t}=\sigma\:({W}_{o}\cdot\:[{h}_{t-1},{x}_{t}]+{b}_{o})$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t}={o}_{t}\text{*}\text{t}\text{a}\text{n}\text{h}\left({C}_{t}\right)$$\end{document} In Eqs. (3)-(8), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{t}$$\end{document} represents the forget gate output, which controls whether information is discarded from the cell state. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document} is the sigmoid function, used to convert inputs to probability values between 0 and 1. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{f}$$\end{document} is the weight matrix of the forget gate. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t-1}$$\end{document} is the hidden state of the previous time step. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}$$\end{document} is the input at the current time step. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{f}$$\end{document} is the bias of the forget gate. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{i}_{t}$$\end{document} is the input gate output, controlling the extent to which new information from the input is stored in the cell state. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{i}$$\end{document} is the weight matrix of the input gate. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{C}}_{t}$$\end{document} is the new candidate cell state, processed by the tanh function, used to update the cell state. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{C}$$\end{document} is the weight matrix of the candidate cell state. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{t}$$\end{document} represents the cell state, updated through the forget gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{t}$$\end{document} and the input gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{i}_{t}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{o}_{t}$$\end{document} is the output gate output, controlling the extent to which the cell state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{t}$$\end{document} is passed to the hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{t}$$\end{document}. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{o}$$\end{document} is the weight matrix of the output gate. GANs are employed to generate synthetic agricultural data for training deep learning models, including crop images and soil data. GANs consist of two main components: a Generator and a Discriminator31. The loss functions for both the Generator and the Discriminator are defined by 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}$$\:\underset{G}{min}\:\underset{D}{max}\:V(D,G)={\mathbb{E}}_{x\sim\:{p}_{data}\left(x\right)}[\text{l}\text{o}\text{g}D(x\left)\right]+{\mathbb{E}}_{z\sim\:{p}_{z}\left(z\right)}[\text{l}\text{o}\text{g}(1-D\left(G\right(z\left)\right)\left)\right]$$\end{document} In Eq. (9), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:V(D,G)$$\end{document} is the objective function of the GANs, measuring the adversarial relationship between the Generator (G) and the Discriminator (D). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbb{E}}_{x\sim\:{p}_{data}\left(x\right)}[\text{l}\text{o}\text{g}D(x\left)\right]$$\end{document} represents the log probability of real data x, used for optimizing the Discriminator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbb{E}}_{z\sim\:{p}_{z}\left(z\right)}[\text{l}\text{o}\text{g}(1-D\left(G\right(z\left)\right)\left)\right]$$\end{document} represents the log probability of generated data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:G\left(z\right)$$\end{document}, used for optimizing the Generator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{data}\left(x\right)$$\end{document} is the distribution of real data, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{z}\left(z\right)$$\end{document} is the distribution of random noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z$$\end{document}. By iteratively training the Generator and the Discriminator, the GANs eventually generates realistic agricultural data, enhancing the effectiveness of model training. Given the constraints on computational resources in practical agricultural applications, this study incorporates several lightweight design features in the deep learning model, such as reducing the number of convolutional layers and minimizing the parameter count32,33. This design approach significantly lowers computational costs while maintaining high accuracy, ensuring deployability on edge devices and in low-resource computing environments. This aspect is particularly crucial in agricultural production scenarios, as it guarantees the model’s efficiency and practicality. In the task of pest and disease detection, the interpretability of the model is essential for agricultural practitioners to understand and trust its outcomes. Although this study does not present specific visualizations, future work will plan to utilize visualization techniques such as Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME) to analyze the model’s decision-making basis34,35. These visualization tools can highlight specific regions of images, illustrating which features contributed to pest and disease classification decisions. Such interpretability can help farmers and agricultural decision-makers better understand the model’s outputs, thereby enhancing the credibility and practicality of pest and disease detection. By integrating these visualization techniques, the model’s application extends beyond accuracy enhancement, providing users with a clear decision-making pathway and thus fostering greater trust in intelligent agricultural systems. Optimization algorithm design To enhance the model’s global search capability and facilitate rapid convergence, this study employs a hybrid optimization method that combines GAs and particle swarm optimization36,37. GA simulates biological evolution by optimizing individual fitness through selection, crossover, and mutation, with the fitness function defined as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(x\right)=\frac{1}{1+f\left(x\right)}$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left(x\right)$$\end{document} is the objective function and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(x\right)$$\end{document}is the fitness function. This study selects GAs and particle swarm optimization as the core combination of optimization techniques, primarily due to their maturity and widespread application in addressing global search and rapid convergence issues. Although novel optimization methods, such as the Crayfish Optimization Algorithm and the Reptile Search Algorithm, have emerged in recent years, their empirical applications in agricultural industry optimization remain limited. GA is renowned for its excellent global search capability, effectively avoiding local optima, while particle swarm optimization ‘s rapid convergence characteristic enables it to quickly locate near-optimal solutions. According to the No Free Lunch Theorem, no single optimization algorithm performs optimally across all problems38. Therefore, the choice to combine GA and particle swarm optimization is grounded in their successful applications to similar problems and their reliability in addressing practical agricultural issues. Particle swarm optimization simulates the foraging behavior of bird flocks, finding the optimal solution through information sharing among individuals39. The velocity and position update formulas for particles are given by Eqs. (10) and (11):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{i}(t+1)=\omega\:{v}_{i}\left(t\right)+{c}_{1}{r}_{1}({p}_{i}-{x}_{i}(t\left)\right)+{c}_{2}{r}_{2}(g-{x}_{i}(t\left)\right)$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}(t+1)={x}_{i}\left(t\right)+{v}_{i}(t+1)$$\end{document} In Eqs. (10) and (11), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\omega\:$$\end{document} is the inertia weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{2}$$\end{document} are acceleration constants, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{2}$$\end{document} is a random number, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}$$\end{document} is the individual best position, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:g$$\end{document} is the global best position. This study employs the combination of GA and particle swarm optimization primarily for the parameter tuning of models, particularly in the hyperparameter optimization processes of CNNs and RNNs. Through meta-heuristic algorithms, key parameters such as kernel size, the number of hidden layer units, and learning rate are searched and optimized to enhance the model’s global search capability and local optimization accuracy. This process plays a critical role in training models for crop yield prediction and pest detection, effectively improving model performance. Due to the inherent randomness of meta-heuristic algorithms, this study conducts 30 independent runs for each optimization experiment to ensure the stability and reproducibility of the results, with the average value taken as the final outcome. Multiple runs help reduce random error and ensure the reliability of the proposed method in practical applications. Adaptive learning algorithms and real-time optimization algorithms Adaptive learning algorithms are designed to enable models to automatically adjust their parameters based on varying environments and conditions, thereby enhancing the model’s adaptability and robustness40,41. For example, crop yield predictions under different climatic conditions require the continuous updating of model parameters through online learning methods. The online gradient descent algorithm is expressed by Eq. (12):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}={\theta\:}_{t}-{\eta\:}_{t}\nabla\:L({\theta\:}_{t};{x}_{t},{y}_{t})$$\end{document} In Eq. (12), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}$$\end{document} represents the model parameters at the t-th step, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\eta\:}_{t}$$\end{document} is the learning rate, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L$$\end{document} is the loss function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{t}$$\end{document} are the input data and true labels, respectively. Transfer learning methods reduce the dependency on large-scale labeled data by transferring knowledge from a pre-trained model to a new task42,43. The fine-tuning process of transfer learning 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}$$\:\mathcal{L}={\mathcal{L}}_{\text{t}\text{a}\text{s}\text{k}}+\lambda\:{\mathcal{L}}_{\text{p}\text{r}\text{e}}$$\end{document} In Eq. (13), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{\text{t}\text{a}\text{s}\text{k}}$$\end{document} is the loss function for the specific task, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{\text{p}\text{r}\text{e}}$$\end{document} is the loss function for the pre-trained model, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} is the balancing coefficient. Real-time optimization algorithms receive real-time data streams and instantly compute the optimal strategy. Real-time optimization based on gradient descent is shown in Eq. (14)44:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t+1}={x}_{t}-\eta\:\nabla\:f\left({x}_{t}\right)$$\end{document} In Eq. (14), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{t}$$\end{document} represents the decision variables at step t, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} is the learning rate, and f is the objective function. Real-time feedback mechanisms are designed to optimize the decision-making process by rapidly responding to external environmental changes and model prediction errors45,46. These mechanisms employ technologies such as sensor networks and wireless communication47,48to collect real-time data on agricultural production and relay it to optimization models for necessary adjustments. The underlying design concept involves gathering data on crop growth status, meteorological conditions, soil moisture, and other relevant factors through sensor networks. This information is transmitted to a central control system using wireless communication technologies49,50. Upon receipt of the data, the central control system conducts real-time analysis and formulates optimization strategies. These strategies are subsequently communicated to agricultural production equipment, such as irrigation and fertilization systems, allowing for real-time adjustments to production operations51,52. The real-time feedback control is represented by Eq. (15)53.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:u\left(t\right)={K}_{p}\left(r\right(t)-y(t\left)\right)+{K}_{i}{\int\:}_{0}^{t}\:\left(r\right(\tau\:)-y(\tau\:\left)\right)d\tau\:+{K}_{d}\frac{d}{dt}\left(r\right(t)-y(t\left)\right)$$\end{document} In Eq. (15), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:u\left(t\right)$$\end{document} represents the control variable, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:r\left(t\right)$$\end{document} denotes the reference signal, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:y\left(t\right)$$\end{document} is the output signal, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{p}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{i}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{d}$$\end{document} are the proportional, integral, and derivative coefficients, respectively. The pseudo code of the above algorithm is as follows: Pseudo code of adaptive learning algorithm: Algorithm: Adaptive Learning Algorithm Input: Initial model parameters θ, learning rate η, threshold T. Repeat until convergence: For each training sample (x_t, y_t): a. Compute the model output: y_pred = Model(x_t, θ). b. Compute loss: L = Loss(y_t, y_pred). c. Update parameters using gradient descent: θ = θ - η * ∇L. d. Adjust learning rate η based on performance. e. If validation loss stops improving for T epochs: Stop training early to avoid overfitting. Output: Optimized model parameters θ. Pseudo code of the hybrid algorithm of genetic algorithm and particle swarm optimization: Algorithm: Hybrid GA-PSO Optimization. Input: Population size N, max generations G, crossover rate C, mutation rate M. Initialize population with random solutions. For generation = 1 to G do: 1. Evaluate fitness of each individual in the population. 2. Select individuals based on fitness (selection). 3. Apply crossover to selected individuals with probability C. 4. Apply mutation to individuals with probability M. 5. Update velocity and position for each particle (PSO step): - Update velocity: v_i = w * v_i + c1 * r1 * (p_best - x_i) + c2 * r2 * (g_best - x_i). - Update position: x_i = x_i + v_i. 6. Replace the worst individuals with new solutions if fitness improves Output: Optimal solution. Real-time feedback optimization algorithm pseudo code: Algorithm: Real-Time Feedback Optimization Input: Initial model parameters θ, real-time data stream D, adjustment threshold T While receiving data D do: 1. Collect new data sample (sensor readings, environmental data, etc.) 2. Update model predictions based on new data 3. If error > T: a. Adjust model parameters θ b. Send feedback to control systems (e.g., irrigation, fertilization) Output: Continuously optimized model and control system actions Experimental design and performance evaluation Datasets collection and experimental environment To enhance the generalization capability of the model, this study utilizes a dataset that encompasses crop growth data from various regions, including multiple crop types (such as rice, wheat, and corn) and diverse climatic conditions (e.g., tropical and temperate). The diversity of these datasets allows the proposed model to adapt to a wider range of agricultural scenarios, thereby increasing its practicality. The datasets employed in this study are primarily sourced from several public repositories: the Ministry of Agriculture’s Crop Pest and Disease Image Database, the Food and Agriculture Organization of the United Nations (FAO) Crop Production Dataset, soil and meteorological data from the National Agricultural Informatization Engineering Technology Research Center, and the Crop Dataset from Kaggle. Specifically, the Agricultural Ministry’s Crop Pest and Disease Image Database contains a substantial collection of high-resolution images depicting various crop pests and diseases, each accompanied by detailed labels and classification information. This image data is instrumental in training and validating CNNs for the efficient identification and classification of crop pests and diseases, thus enabling farmers to implement timely preventive measures, enhance production efficiency, and improve crop quality. The Food and Agriculture Organization Crop Production Dataset provides comprehensive information on crop yields, planting areas, and production conditions across multiple global regions. This dataset is utilized to train RNNs and LSTM networks for time series forecasting, such as evaluating the impacts of climate change on crop yields. This approach offers a scientific basis for agricultural production decisions, thereby improving resource utilization efficiency and promoting agricultural sustainability. The soil and meteorological data sourced from the National Agricultural Informatization Engineering Technology Research Center include detailed insights into soil types, soil fertility, and various meteorological conditions (e.g., temperature, precipitation, humidity). This data is leveraged to optimize various facets of agricultural production through deep learning models and hybrid optimization algorithms, facilitating precise monitoring and forecasting of crop growing environments and enhancing the intelligent management of agricultural production. The Crop Dataset from Kaggle encompasses several years of crop growth data, including information on crop types, growth cycles, yields, and the corresponding meteorological and soil conditions. This dataset is utilized to train GANs to generate additional synthetic agricultural data, thereby augmenting the model’s generalization capabilities and demonstrating its significant potential to improve agricultural production efficiency and environmental sustainability through empirical experimentation. To verify the effectiveness of the synthetic data generated by the GAN, this study used Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR) as similarity evaluation metrics54. The comparison results between the generated data and real data on these metrics are presented in Table 1. Table 1Effectiveness evaluation of GAN-generated data.Dataset typeSSIMPSNR (dB)GAN-generated0.9228.5Real data130 Table 1Effectiveness evaluation of GAN-generated data.Dataset typeSSIMPSNR (dB)GAN-generated0.9228.5Real data130 Effectiveness evaluation of GAN-generated data. The results in Table 1 indicate that the GAN-generated data exhibits high similarity in structure and detail compared to the real data, demonstrating its effectiveness for use in model training. The experimental environment is detailed in Table 2. Table 2Software and hardware environment.CategoryNameSpecification/version/descriptionHardware environmentProcessorIntel Core i7-10700 K @ 3.80 GHz (Gigahertz)Memory32 GB (Gigabytes) Double Data Rate (DDR) 4Graphics cardNvidia GeForce Ray Tracing Acceleration (RTX) 3080 TiStorage1 TB (Terabyte) Solid State Drive (SSD)Operating systemWindows 10ServerAmazon Web Services (AWS) Elastic Compute (EC) 2 t3.xlargeGraphics Processing Unit (GPU)Nvidia Tesla V100Software environmentPython3.8.10TensorFlow2.7.0PyTorch1.10.0Keras2.8.0Compute Unified Device Architecture (CUDA)11.5Deep Neural Network library (cuDNN)8.2.1 Table 2Software and hardware environment.CategoryNameSpecification/version/descriptionHardware environmentProcessorIntel Core i7-10700 K @ 3.80 GHz (Gigahertz)Memory32 GB (Gigabytes) Double Data Rate (DDR) 4Graphics cardNvidia GeForce Ray Tracing Acceleration (RTX) 3080 TiStorage1 TB (Terabyte) Solid State Drive (SSD)Operating systemWindows 10ServerAmazon Web Services (AWS) Elastic Compute (EC) 2 t3.xlargeGraphics Processing Unit (GPU)Nvidia Tesla V100Software environmentPython3.8.10TensorFlow2.7.0PyTorch1.10.0Keras2.8.0Compute Unified Device Architecture (CUDA)11.5Deep Neural Network library (cuDNN)8.2.1 Software and hardware environment. Parameters setting During the model training process, parameters are dynamically adjusted in response to real-time feedback and experimental results to optimize model performance. A learning rate decay strategy is implemented, where the learning rate is halved every 10 epochs to mitigate issues of overfitting or underfitting. Model performance is continuously monitored on the validation set, and training is halted early when the error rate on the validation set no longer decreases, thereby preventing overfitting. Batch sizes are adjusted dynamically according to the convergence speed of the model and the available computational resources, aiming to balance training time with model stability. Xavier initialization is employed, selecting suitable weight initialization strategies based on the model’s depth and the characteristics of the activation functions to ensure stable convergence. The parameter settings for the model utilized in this study are summarized in Table 3. Table 3Parameter settings.Model typeParametersValueCNNsConvolution kernel size3 × 3Number of convolution layers4Pooling layer size2 × 2Learning rate0.001Batch size32RNNsNumber of hidden layers3Number of hidden units128Learning rate0.01LSTMInput gate weightInitialize to random valueForget gate weightInitialize to random valueOutput gate weightInitialize to random valueLearning rate0.01GANsNumber of generator layers3Number of discriminator layers3Learning rate (generator)0.0002Learning rate (discriminator)0.0002GAs + particle swarm optimizationPopulation size50Number of iterations100Crossover probability0.8Mutation probability0.1Inertia weight0.5Acceleration constantc1 = 2, c2 = 2 Table 3Parameter settings.Model typeParametersValueCNNsConvolution kernel size3 × 3Number of convolution layers4Pooling layer size2 × 2Learning rate0.001Batch size32RNNsNumber of hidden layers3Number of hidden units128Learning rate0.01LSTMInput gate weightInitialize to random valueForget gate weightInitialize to random valueOutput gate weightInitialize to random valueLearning rate0.01GANsNumber of generator layers3Number of discriminator layers3Learning rate (generator)0.0002Learning rate (discriminator)0.0002GAs + particle swarm optimizationPopulation size50Number of iterations100Crossover probability0.8Mutation probability0.1Inertia weight0.5Acceleration constantc1 = 2, c2 = 2 Parameter settings. Performance evaluation Model performance evaluation This section compares the performance of the proposed method against several benchmark algorithms in terms of accuracy, precision, recall, F1 score, number of parameters (M), training time (hrs), and inference time (ms). The benchmark algorithms include Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Machine (GBM), and eXtreme Gradient Boosting (XGBoost). Additionally, three optimization algorithms—Crayfish Optimization, Reptile Search, and Red Fox Optimizer—are included for comparison. The results are shown in Table 4. Table 4Comparison of model performance.MethodAccuracyPrecisionRecallF1 scoreParameters (M)Training time (hrs)Inference time (ms)The proposed method0.920.910.930.922.14.120SVM0.860.840.860.850.83.818RF0.870.860.880.871.24.225GBM0.880.870.880.881.54.922XGBoost0.890.880.90.891.75.124Crayfish optimization0.890.880.90.891.95.323Reptile search0.870.860.880.871.64.721Red fox optimizer0.850.840.860.851.44.522 Table 4Comparison of model performance.MethodAccuracyPrecisionRecallF1 scoreParameters (M)Training time (hrs)Inference time (ms)The proposed method0.920.910.930.922.14.120SVM0.860.840.860.850.83.818RF0.870.860.880.871.24.225GBM0.880.870.880.881.54.922XGBoost0.890.880.90.891.75.124Crayfish optimization0.890.880.90.891.95.323Reptile search0.870.860.880.871.64.721Red fox optimizer0.850.840.860.851.44.522 Comparison of model performance. In Table 4, the proposed method achieves an accuracy of 0.92, precision of 0.91, recall of 0.93, and an F1 score of 0.92, all of which surpass the performance of SVM, RF, GBM, and XGBoost. This indicates that the proposed method exhibits the best overall performance in classification tasks, particularly excelling in precision and recall. In contrast, SVM achieves an accuracy of only 0.86, RF achieves 0.87, while GBM and XGBoost yield slightly higher performance, yet still fall short of the proposed method. Thus, the proposed method demonstrates superior predictive accuracy and consistency in comprehensive performance, making it suitable for practical applications in smart agriculture systems. The classification performance of the three newly added algorithms—Crayfish Optimization, Reptile Search, and Red Fox Optimizer—is close to that of XGBoost and GBM, though none surpasses the performance of the proposed method. The proposed method has a moderate parameter count of 2.1 M, with a relatively short training time of 4.1 h and inference time of 20 ms, indicating its suitability for deployment in practical agricultural applications. In contrast, Crayfish Optimization requires 5.3 h for training and 23 ms for inference; while its accuracy approaches that of the proposed method, its computational cost and efficiency are lower. Among all benchmarked algorithms, the proposed method achieves the best inference time at only 20 ms. Although SVM has a slightly faster inference time of 18 ms, its lower classification performance limits its applicability in complex agricultural scenarios. To determine whether the performance differences among the various algorithms on these metrics are statistically significant, a T-test is conducted, with results presented in Table 5: Table 5Results of the T-test.MetricCompared modelsMean differenceT valueDegrees of freedomp valueStatistical significanceAccuracyProposed Method vs. SVM0.064.2180.0002SignificantAccuracyProposed Method vs. RF0.053.9180.0004SignificantAccuracyProposed Method vs. Crayfish0.032.6180.012SignificantAccuracyProposed Method vs. Reptile Search0.053.7180.001SignificantAccuracyProposed Method vs. Red Fox0.074.3180.0003SignificantPrecisionProposed Method vs. GBM0.043180.005SignificantPrecisionProposed Method vs. Crayfish0.032.5180.015SignificantRecallProposed Method vs. XGBoost0.032.8180.008SignificantRecallProposed Method vs. Reptile Search0.053.4180.002SignificantRecallProposed Method vs. Red Fox0.074.1180.0005SignificantF1 scoreProposed Method vs. SVM0.074.5180.0001SignificantF1 scoreProposed Method vs. Crayfish0.032.7180.01Significant Table 5Results of the T-test.MetricCompared modelsMean differenceT valueDegrees of freedomp valueStatistical significanceAccuracyProposed Method vs. SVM0.064.2180.0002SignificantAccuracyProposed Method vs. RF0.053.9180.0004SignificantAccuracyProposed Method vs. Crayfish0.032.6180.012SignificantAccuracyProposed Method vs. Reptile Search0.053.7180.001SignificantAccuracyProposed Method vs. Red Fox0.074.3180.0003SignificantPrecisionProposed Method vs. GBM0.043180.005SignificantPrecisionProposed Method vs. Crayfish0.032.5180.015SignificantRecallProposed Method vs. XGBoost0.032.8180.008SignificantRecallProposed Method vs. Reptile Search0.053.4180.002SignificantRecallProposed Method vs. Red Fox0.074.1180.0005SignificantF1 scoreProposed Method vs. SVM0.074.5180.0001SignificantF1 scoreProposed Method vs. Crayfish0.032.7180.01Significant Results of the T-test. Table 5 provides the results of the T-tests comparing the proposed method with various benchmark models, including SVM, RF, GBM, XGBoost, Crayfish Optimization, Reptile Search, and Red Fox Optimizer, across different performance metrics. The T-test results yield p-values, which indicate statistical significance. Generally, a p-value less than 0.05 (as observed here) suggests that the improvements of the proposed method on the corresponding metrics are statistically significant. For accuracy, the mean difference between the proposed method and SVM is 0.06, with a T-value of 4.20, degrees of freedom of 18, and a p-value of 0.0002, demonstrating significance at the p < 0.05 level. To evaluate the system’s performance in regression tasks, the mean absolute error (MAE) and mean squared error (MSE) are employed as metrics. Figure 3 illustrates the performance of the proposed method relative to these comparison algorithms across the same metrics. Fig. 3Performance comparison in regression tasks. Performance comparison in regression tasks. In Fig. 3, the proposed method demonstrates superior performance in the regression task compared to the other benchmark algorithms. Specifically, the Mean Absolute Error (MAE) for the proposed method is 1.25, and the MSE is 2.35, both of which are lower than the corresponding values for SVM, RF, GBM, XGBoost, Crayfish Optimization, Reptile Search, and Red Fox Optimizer. These results indicate that the proposed method exhibits enhanced predictive accuracy with smaller error margins. When evaluating the performance of metaheuristic algorithms, it is essential to include metrics such as the best, worst, mean, median, standard deviation, and variance scores of the objective function. Table 6Objective function scores.AlgorithmBestWorstMeanMedianStdVarThe proposed method0.920.880.900.910.020.0004SVM0.860.830.850.850.0150.0002RF0.880.850.870.870.010.0001GBM0.890.860.880.880.010.0001XGBoost0.910.890.900.900.010.0001Crayfish optimization0.910.870.890.890.0150.0002Reptile search0.880.850.860.860.0150.0002Red fox optimizer0.870.840.850.850.0150.0002 Table 6Objective function scores.AlgorithmBestWorstMeanMedianStdVarThe proposed method0.920.880.900.910.020.0004SVM0.860.830.850.850.0150.0002RF0.880.850.870.870.010.0001GBM0.890.860.880.880.010.0001XGBoost0.910.890.900.900.010.0001Crayfish optimization0.910.870.890.890.0150.0002Reptile search0.880.850.860.860.0150.0002Red fox optimizer0.870.840.850.850.0150.0002 Objective function scores. In Table 6, the proposed method exhibits outstanding performance in terms of objective function scores, achieving a best score of 0.92 and a worst score of 0.88, with a mean of 0.90, indicating overall stability in performance. The median score is 0.91, suggesting that most of the run results are close to the best score. Furthermore, the method demonstrates low variability, as indicated by a standard deviation of 0.02 and a variance of 0.0004. In contrast, the SVM shows a best score of 0.86 and a worst score of 0.83, reflecting poorer performance with greater variability (standard deviation of 0.015). The performance of XGBoost is also comparable to that of the proposed method, though it exhibits slightly lower variability (standard deviation of 0.01). Overall, the proposed method outperforms the other comparative algorithms in both accuracy and stability. In contrast, Crayfish Optimization achieves a mean score of 0.89, which is close to that of the proposed method, though it exhibits slightly higher volatility with a standard deviation of 0.015. Additionally, Reptile Search and Red Fox Optimizer yield lower mean scores of 0.86 and 0.85, respectively, with standard deviations of 0.015, indicating relatively higher fluctuations in performance. In the performance evaluation of metaheuristic algorithms, the box plot shown in Fig. 4 illustrates the performance distribution of each model over multiple independent runs. The figure includes the performance results of eight models: the proposed method, SVM, RF, GBM, XGBoost, Crayfish Optimization, Reptile Search, and Red Fox Optimizer. Fig. 4Performance box plots of different models. Performance box plots of different models. In Fig. 4, the proposed method demonstrates the highest median score and exhibits a smaller variance, indicating greater stability and consistency in its performance. In contrast, the SVM shows a more dispersed distribution, reflecting significant fluctuations in performance across different runs, while the performances of RF, GBM, and XGBoost fall between these two extremes. Overall, the proposed method shows superior stability and consistency across multiple independent runs. The performance scores of the Crayfish Optimization and XGBoost methods are close to those of the proposed method; however, the proposed method still demonstrates a slight advantage in terms of median performance and consistency. This consistency is beneficial for applications requiring reliable performance, such as agricultural forecasting. The convergence data of different algorithms over several iterations is presented, allowing the optimization processes of each model to be clearly observed. The convergence rates of the various algorithms across multiple iterations are illustrated in Fig. 5: Fig. 5Convergence curve. Convergence curve. The data in Fig. 5 is shown in Table 7: Table 7Model convergence data.IterationsProposed methodSVMRFGBMXGBoostCrayfish optimizationReptile searchRed fox optimizer06.57.26.876.96.877.1503.25.45.65.24.95.15.35.51002.74.94.84.64.44.54.74.81501.94.44.343.83.94.14.22001.53.93.83.73.43.53.73.8 Table 7Model convergence data.IterationsProposed methodSVMRFGBMXGBoostCrayfish optimizationReptile searchRed fox optimizer06.57.26.876.96.877.1503.25.45.65.24.95.15.35.51002.74.94.84.64.44.54.74.81501.94.44.343.83.94.14.22001.53.93.83.73.43.53.73.8 Model convergence data. In Fig. 5; Table 7, the proposed optimization method demonstrates convergence to a relatively low error (MSE = 1.9) after approximately 150 iterations, which is faster compared to other models such as SVM, RF, and GBM. The convergence speed is notably quicker, and the error is lower. Although XGBoost also performs relatively well, it still falls short of the performance exhibited by the proposed method. This observation indicates that the proposed method offers a faster convergence rate and higher computational efficiency, making it suitable for processing and predicting large-scale agricultural data. In contrast, Crayfish Optimization and XGBoost exhibit relatively fast convergence rates, achieving MSE of 3.5 and 3.4, respectively, after 200 iterations, although they remain slightly inferior to the proposed method. Conversely, Reptile Search and Red Fox Optimizer show slower convergence rates and higher convergence errors, indicating their lower efficiency for this specific task compared to the proposed method. Efficiency evaluation of models The performance regarding computational costs of the different models is presented in Fig. 6: Fig. 6Comparison of computational costs. Comparison of computational costs. The data of Fig. 6 is shown in Table 8: Table 8Comparison data of calculation costs.MethodTraining time (hours)Memory usage (GB)GPU usage rateThe proposed method4.180.8SVM3.860.45RF4.270.5GBM4.980.6XGBoost5.180.7Crayfish optimization4.370.55Reptile search4.680.6Red fox optimizer4.77.50.65 Table 8Comparison data of calculation costs.MethodTraining time (hours)Memory usage (GB)GPU usage rateThe proposed method4.180.8SVM3.860.45RF4.270.5GBM4.980.6XGBoost5.180.7Crayfish optimization4.370.55Reptile search4.680.6Red fox optimizer4.77.50.65 Comparison data of calculation costs. In Table 8, the proposed method demonstrates a distinct advantage in terms of computational cost. Specifically, the training time for the proposed method is 4.1 h, with a memory usage of 8 GB and a GPU utilization rate of 0.8. In contrast, the training time for the SVM is shorter at 3.8 h, while XGBoost has the longest training time at 5.1 h. Scalability in large-scale applications is another critical metric for evaluating model performance. The proposed method is evaluated as a whole and compared with other models, focusing on key evaluation metrics such as data processing speed, model training stability, and scalability (rated on a scale of 1 to 10), as illustrated in Fig. 7. Fig. 7Scalability comparison. Scalability comparison. In Fig. 7, the proposed method demonstrates superior performance in data processing speed and model training stability, achieving a high scalability score. Specifically, the data processing speed reaches 1,550 items per second, with model training stability recorded at 93%, and a scalability rating of 9 points. This method outperforms the comparative algorithms in both data processing speed and model training stability. Comparison of practical application effects of the model To assess the effectiveness of the model in predicting crop yields, the proposed method is compared with other models using a dataset that includes multiple years of climate data, soil data, and crop growth data. The proposed method employs deep learning models to predict multi-dimensional time series data, enhancing prediction accuracy through a hybrid optimization algorithm. Figure 8 illustrates the performance of the various models in crop yield prediction: Fig. 8Comparison of crop yield prediction. Comparison of crop yield prediction. The data in Fig. 8 are shown in Table 9: Table 9Comparative data on crop yield prediction.MethodMAEMSE R 2 The proposed method210.0450.93SVM320.10.87RF280.07840.89GBM260.06760.9XGBoost230.05290.92Crayfish optimization240.0590.90Reptile search270.0730.89Red Fox optimizer290.0820.88 Table 9Comparative data on crop yield prediction.MethodMAEMSE R 2 The proposed method210.0450.93SVM320.10.87RF280.07840.89GBM260.06760.9XGBoost230.05290.92Crayfish optimization240.0590.90Reptile search270.0730.89Red Fox optimizer290.0820.88 Comparative data on crop yield prediction. In Fig. 8; Table 9, the proposed method exhibits superior performance in terms of MAE and MSE metrics, achieving the highest R² value of 0.93. This indicates enhanced accuracy and stability in crop yield prediction when compared to other models. The data demonstrates that the proposed method effectively manages multi-dimensional time series data and improves prediction performance through the application of optimization algorithms. The evaluation of pest and disease detection is conducted using a dataset comprising large-scale crop images. The proposed method employs CNNs for image recognition and classification. Additionally, it utilizes GANs to generate synthetic data, thereby enhancing the model’s generalization ability. The results are presented in Table 10: Table 10Comparison of pest and disease detection.MethodAccuracyPrecisionRecallF1 scoreThe proposed method0.9750.9680.9720.97SVM0.9040.8970.8930.9RF0.9210.9130.9160.92GBM0.9420.9380.9350.94XGBoost0.9580.9520.9490.95Crayfish optimization0.9530.9470.9440.94Reptile search0.9330.9260.9240.92Red fox optimizer0.9170.9120.910.91 Table 10Comparison of pest and disease detection.MethodAccuracyPrecisionRecallF1 scoreThe proposed method0.9750.9680.9720.97SVM0.9040.8970.8930.9RF0.9210.9130.9160.92GBM0.9420.9380.9350.94XGBoost0.9580.9520.9490.95Crayfish optimization0.9530.9470.9440.94Reptile search0.9330.9260.9240.92Red fox optimizer0.9170.9120.910.91 Comparison of pest and disease detection. In Table 10, the proposed method demonstrates superior performance in pest and disease detection, achieving an accuracy of 97.5%, precision of 96.8%, recall of 97.2%, and an F1 score of 0.97. These results highlight its effectiveness in accurately identifying pests and diseases. By employing CNNs for image recognition and integrating GANs to generate additional data, the proposed method significantly enhances the model’s generalization ability and detection accuracy. Furthermore, this study evaluates the impact of intelligent transformation on agricultural productivity and economic efficiency, with the results presented in Table 11: Table 11Evaluation of the impact of intelligent transformation on agricultural productivity and economic efficiency.IndicatorBefore smart transitionAfter smart transitionProduction cost (per acre)$500$450Pest/disease loss10%3%Crop yield (tons/acre)3.54.2 Table 11Evaluation of the impact of intelligent transformation on agricultural productivity and economic efficiency.IndicatorBefore smart transitionAfter smart transitionProduction cost (per acre)$500$450Pest/disease loss10%3%Crop yield (tons/acre)3.54.2 Evaluation of the impact of intelligent transformation on agricultural productivity and economic efficiency. In Table 11, the impact of intelligent transformation on agricultural productivity and economic efficiency is significant. Firstly, production costs have decreased from $500 per acre to $450, indicating that smart agricultural technologies effectively lower costs by enhancing resource utilization efficiency. Secondly, the pest and disease loss rate has been reduced from 10 to 3%, demonstrating the excellent performance of intelligent detection technologies in promptly identifying and controlling pests and diseases, thereby significantly reducing crop losses. Finally, crop yield has increased from 3.5 tons per acre to 4.2 tons, highlighting the positive role of intelligent technologies in optimizing agricultural management and improving production efficiency. These improvements illustrate the enhanced economic benefits and increased productivity resulting from smart agricultural technologies. Discussion Compared to traditional methods such as SVM, RF, and GBM, the proposed method demonstrates superior capabilities in handling large-scale and diverse agricultural data, resulting in higher prediction accuracy and stability. This outcome is primarily attributed to the effective tuning of model parameters through the hybrid optimization algorithm, enabling the model to more accurately capture the underlying patterns and relationships within complex tasks, thereby enhancing predictive accuracy. In the study by Hadipour-Rokni et al. (2023), which utilized only CNNs for pest and disease identification, although significant improvements in generalization and accuracy were achieved for specific tasks. The experiment lacked a comprehensive consideration of agricultural industry structure optimization by integrating CNNs with GANs for synthetic data generation55. The data generated by GANs closely resembles real data in terms of SSIM and PSNR, validating its effectiveness in large-scale agricultural image data. This data generation strategy addresses the shortcomings of traditional machine learning methods, further enhancing the robustness and detection accuracy of pest and disease identification models. In the realm of climate change prediction, Park et al. (2023) employed RNNs and LSTM networks for time series data analysis, achieving favorable results; however, they did not incorporate optimization algorithms for further model refinement56. This experiment significantly improves the data processing capabilities of RNNs and LSTMs under varying environmental conditions through hybrid optimization. For instance, in the crop yield prediction task, the proposed method achieved a MSE of 2.35, significantly outperforming benchmark algorithms. Continuous tuning of the model via the hybrid optimization algorithm demonstrated strong adaptability in time series forecasting tasks, maintaining high prediction accuracy and stability despite fluctuations in climate and environmental conditions. While metaheuristic algorithms such as Crayfish Optimization, Reptile Search, and Red Fox Optimizer approach the performance of the proposed method on certain metrics, they exhibit slight deficiencies in computational efficiency and stability. The proposed method outperforms these algorithms in training time (4.1 h) and inference time (20 ms), showcasing its efficiency and adaptability for practical deployment. Additionally, the low variance observed across multiple independent runs indicates more stable and consistent performance. These findings highlight the unique advantages of the proposed method in integrating hybrid optimization with deep learning, making it suitable for addressing complex agricultural tasks in real-world applications. Regarding the impact of smart transformation on agricultural productivity and economic benefits, the experimental results indicate that more accurate yield predictions and pest detection can effectively reduce production costs (from $500 to $450 per acre) and significantly decrease losses due to pests and diseases (from 10 to 3%). This improvement is primarily driven by the deep learning model’s effective handling of multidimensional data in complex environments. The integration of GAN-generated data with the hybrid optimization algorithm enables the model to better adapt to dynamic agricultural settings, thereby supporting enhanced production efficiency and resource optimization. These results not only validate the practical application value of the proposed method but also lay a solid foundation for future advancements in the smart transformation of agriculture. Conclusion Research contribution This study addresses the core challenges of optimizing agricultural industrial structures and facilitating the smart transformation of agriculture by proposing an intelligent agricultural system that integrates deep learning and optimization technologies. By employing CNNs for crop pest and disease recognition and classification, RNNs and LSTM networks for predicting climate change and crop growth cycles, GANs for synthesizing agricultural data, and a hybrid optimization method that combines GAs and particle swarm optimization, the proposed system successfully enhances the precision and efficiency of agricultural production. The model incorporates adaptive learning and a real-time optimization feedback mechanism, thereby improving decision-making quality under varying environmental conditions. This system excels in processing and optimizing multidimensional agricultural data, offering new pathways for the intelligent transformation of agriculture. Future works and research limitations Despite these significant achievements, this study acknowledges certain limitations. First, although the model demonstrates commendable performance, its generalization ability across different agricultural scenarios still requires further validation. Future work will integrate the optimization system with real-time monitoring technologies to create a dynamic feedback mechanism, allowing adjustments to agricultural management strategies based on real-time data and enhancing the model’s practical applicability. Second, this study has yet to present specific visualizations of the model’s decision-making processes. Future plans include employing visualization techniques such as Grad-CAM and LIME to analyze the model’s decision basis, thereby improving its interpretability and transparency. Additionally, while deep learning models exhibit superior performance, their computational costs and resource requirements remain a concern. Future research will explore the development of lightweight models to accommodate large-scale application scenarios.