LDDHybridNet: an ROI-aware CNN–LSTM hybrid framework for accurate and early leaf disease detection in precision agriculture Authors: Ch Santoshi, Suresh Kumar Mandala, L Raghavendar Raju Published: 2026 ABSTRACT AbstractEarly and accurate detection of plant leaf diseases is an essential requirement for precision agriculture, given their severe impact on global food security. While much has been done recently, many deep learning–based approaches will still fail in real-world tests because of challenges such as background clutter, differences in illumination, occlusion, or the fact that visual symptoms for these diseases can be very subtle early on. Traditional CNN- and Transformer-based architectures generally lack accurate lesion localisation and interpretability, hindering their practical deployment in agricultural decision-support tools. To address these issues, we present LDDHybridNet, a region-based, explanation-friendly deep learning framework that can identify leaf disease at an early, accurate stage. It then applies preprocessing steps guided by ROI, based on leaf segmentation from the U-Net, followed by a compact CNN-based spatial feature-extraction framework. We arrange spatial feature embeddings extracted from lesion regions into an ordered sequence and employ a Bi-LSTM with attention to model structured contextual dependencies, allowing progression-aware feature learning without requiring actual temporal image sequences. Lastly, Grad-CAM–based post-hoc explainability is employed to interpret model decisions, enabling transparent visualisation of disease-relevant regions. We conduct extensive experiments on the PlantVillage benchmark and the FieldPlant dataset and show that LDDHybridNet consistently outperforms representative CNN, transformer, and hybrid baselines across multiple evaluation metrics. Although the near-ceiling performance on PlantVillage reveals the dataset’s artificial nature, the proposed framework achieves 95.37% accuracy under real-world field conditions and 92.84% on weak-lesion early-stage samples, demonstrating the method’s robustness and early-stage detection potential. The performance boosts are statistically significant (P < 0.01). In general, LDDHybridNet is an interpretable and robust deep learning framework for leaf disease detection, which can support data-driven crop protection and precision agriculture applications.Keywords: Leaf Disease Detection, Precision Agriculture, CNN–LSTM Hybrid Model, Region of Interest Extraction, Explainable Deep LearningSubject terms: Computational biology and bioinformatics, Engineering, Mathematics and computing Introduction Plant leaf disease epidemics threaten regional and global crop yields, food security, and agroecosystem sustainability—introduction (max 15 Bullet points). Early and precise detection of crop disease is essential to avert crop losses and facilitate timely interventions. Traditional manual inspection methods are subjective, time-consuming, and non-scalable. Automatic plant disease diagnosis using deep learning technologies is a new research hot spot, with potential significance for ensuring the growing food demand while agricultural land is limited1,2. However, research indicates that these models do not generalise well in data-invariant field settings, which are characterised by variable ambient lighting, cluttered and occluded background scenes that change over time, and shifts between laboratory and field imagery domains3,4. Hybrid and transformer-based models have attempted to tackle multi-scale lesion patterns and global context5,6, but they remain primitive for early-symptom detection, lacking a specialised module to segregate disease-specific regions. Furthermore, current studies are challenging to interpret and offer little guidance on the biological logic underpinning model predictions7,8. These limitations indicate the possibility of a novel deep learning framework with robust ROI extraction, improved learning of spatial–temporal features, and, most importantly, interpretable decisions. This paper aims to develop a unified architecture for precise early leaf disease identification (LDI) with explainability across controlled & field conditions. LDDHybridNet performs: (i) the two-stage ROI pipeline (U-Net + YOLO) that uses input ROI image patches to guide feature extraction while remaining fully connected using a low-dimensional global representation offered through vision transformer (ViT) architecture; (ii) multi-scale ASPP PE that combines local and global representations with ViT-based global representation that captures more global context while modeling many channels; and (iii) the attention mechanism through both the SE and CBAM attention mechanisms as feature selection approaches for advanced feature selection, and (iv) end-to-end tracking of the lesion through the temporal consistent Bi-LSTM having the global temporal modeling ability with the attention mechanism as a light-weight learner to model the query representation of the temporal context for the lesion progression modeling. This work provides the following novelties: (i) a precise ROI-isolating method operating in a two-stage fashion; (ii) a unified spatial–temporal hybrid architecture capable of differentiating and capturing high-resolution, fine-grained and evolving disease cues; (iii) improvement on the early-stage detection capability; and (iv) biologically meaningful interpretability represented in a new layer through validating Grad-CAM. Approaches to recent precision agriculture trends have been centred more on deployment-oriented, near–real-time visual analytics rather than on aspects of early-phase plant monitoring, especially for early-stage plants in resource-constrained settings. Deep learning solutions compatible with Edge have already proven practical for plant-based analysis, achieving lower latency and reduced computational load for real-time seed detection and germination evaluation on embedded platforms9. More specifically, we may even find lightweight semantic segmentation networks for agricultural images10. Our results underscore the need for region-wise processing, a small architecture, and post-processing for feature extraction of real-life farm scenarios. Informed by these advancements, we thus take a deployment-aware design approach in this work, aiming for robustness, interpretability, and early-disease sensitivity, rather than necessarily validated real-time performance. This paper has the following contributions. It exhibits substantial end-to-end benchmark superiority over state-of-the-art CNN, transformer and hybrid models on the PlantVillage and FieldPlant datasets;2) it shows a good generalisation under difficult field conditions;3) it discerns each element via ablation and statistical significance analyses;4) it offers explainability-driven visual reasoning, which can potentially guide practical agronomic decision-making. The remainder of this paper is structured as follows: Sect. 2 presents a comprehensive review of the literature, trends, and research gaps. In Sect. 3, we introduce our method, LDDHybridNet, including ROI extraction, architecture, and training details. In Sect. 4, we present our experimental results, quantitative evaluations, and analyses of early-stage detection and interpretability. In Sect. 5, I will present the study’s excerpts and limitations. Finally, Sect. 6 concludes the paper and presents future research directions. Related work It summarises CNN-based, Hybrid, ROI-aware, Explainable, and deployment-oriented techniques for reviewing recent developments in deep learning-based leaf disease detection and identifying their limitations, which motivate our proposed LDDHybridNet framework. Deep learning paradigms for leaf disease detection Deep learning has revolutionised automated leaf disease detection by moving away from handcrafted features and integrating feature learning into end-to-end training on raw images. CNN-based models were found to be superior in terms of accuracy, robustness, and scalability over traditional machine learning methods for plant disease classification in a few early comparative studies11,12. Later review papers systematically chronicled this evolution and identified CNN-based pipelines as the method of choice across a variety of crops and imaging conditions13–17. These studies highlighted the capacity of deep architectures to learn hierarchical representations that relate to the proper colour, texture, and morphology of lesions and are essential for distinguishing diseases. Nonetheless, they have also pointed out recurring difficulties with background clutter, variability in disease stage, and limited generalisation across crops, which have led to persistent motivations for architectural improvements18–20. CNN architectures and transfer learning in crop-specific disease detection A significant amount of research has been directed toward developing CNN networks customised for different crops and disease classes. VGG-based models21, EfficientNet variants5,22,23, and compound-scaled CNNs24 have shown strong performance with light-weight and scalable networks. Transfer learning13,25 is a common approach for small annotated datasets8 that allows the models to generalize faster and better for rice26,27, apple22, grape24,28, maize29, jasmine30, and citrus2,6 and tomato crops31. Although they can yield high classification performance under controlled conditions, such methods generally process whole-leaf images, which (I) often dilute disease information via spatial averaging and (II) also hamper early-stage symptom detection, especially in complex field environments32–34. Hybrid deep learning models with sequential and contextual learning To improve representational power over spatial features, hybrid deep learning models that combine CNNs with sequence learners such as LSTM, BiLSTM, and GRU have gained popularity. Such architectures harness contextual and temporal dependencies, which are vital for disease progression analysis and challenging background modelling. Prommakhot et al. Some common approaches include CNN feature extractors + BiLSTM and CNN feature extractors + transformer-based sequential learning to enhance discriminability1, and CNN–BiLSTM fusion, which improves robustness for real-time tomato disease detection3. Contextual hybrid models that measure spatio–temporal imagery and multispectral data have also been implemented for agricultural monitoring35,36. Overall, these studies show that CNN–RNN hybrids work better than CNNs alone. Still, most existing works do not provide explicit spatial attention mechanisms to extract disease-specific regions for sequential learning. ROI-Aware, segmentation, and localised disease analysis Over the last couple of years, activities have focused on ROI (Region of Interest) learning and localised disease-detection approaches to enhance the precision of early diagnosis. In this regard, patch-based and lesion-centric strategies were found to improve crop generalisation by minimising reliance on global leaf structure37,38. Then, this localisation of disease provides greater interpretability and reliability39. Particularly in the context of detecting tomato diseases, object detection-based models address this problem in two significant aspects: the recognition of small lesions when partially covered and under complex background conditions31. Real-time dataset development and field deployment analyses focused on ROI-guided computation to minimize the inference cost at the expense of accuracy preservation32. However, most ROI-aware methods are spatially static and do not directly model the temporal evolution of the disease. Explainable, multimodal, and deployment-oriented frameworks This indicates the increasing deployment of AI systems in precision agriculture and, hence, an increasing demand for interpretability, contextual awareness, and real-world feasibility. Grad-CAM and attention visualisation and other explainable AI (XAI) techniques have been used to confirm where the model is focused so that it can be trusted5,6, as evidenced by numerous reviews that highlight that XAI is an essential facilitator for adoption4,14,26. Multimodal disease prediction studies have also been conducted in parallel, in which multimodal environmental and Internet of Things (IoT) sensor streams, in addition to image data, were fused to provide disease risk and potentially an early warning system25,40–42. However, citizen science imagery and UAV-based sensing research also identified user-scaling advantages and data-quality challenges7. On the other hand, lightweight and edge-friendly models have been proposed to facilitate the real-time operation of agriculture NPUs optimised detection models8. While these methods improve functionality, they also introduce architectural overhead and are not well-integrated, lacking a standard design for understanding ROI, temporal modelling, and explainability. A critical review of the literature has revealed remarkable advancements in CNN-based, hybrid, ROI-aware, and explainable leaf disease detection systems. Nevertheless, current approaches usually consider these aspects in a decoupled manner. Despite continuous advancements, there remains an evident need for a single framework that simultaneously incorporates ROI-aware spatial learning, temporal modelling of disease progression, and interpretability, whilst being computationally efficient enough for precision agriculture. First, this gap has been the primary motivation for the proposed LDDHybridNet, which locally blends and integrates (after extraction) ROI-guided CNN feature extraction with LSTM-based sequential learning to yield LDD that is accurate, early, and interpretable. In Table 1, a summarised comparison of the existing literature over the last couple of years is provided, focusing on deep learning–based leaf disease detection studies and discussing the core methodologies, strengths, shortcomings, and research gaps that necessitate the proposed LDDHybridNet framework. Table 1Summary of Related Work on Deep Learning–Based Leaf Disease Detection.CategoryRepresentative WorksCore MethodologyKey StrengthsLimitationsIdentified Research GapCNN vs. Traditional ML FoundationsSujatha et al39., Li et al40., Jackulin & Murugavalli13, Doutoum & Tugrul38CNN-based image classification compared with ML feature-based modelsDemonstrated superiority of CNNs in accuracy and robustnessSensitive to background noise and disease stageLack of spatial focus and early-stage disease sensitivityCrop-Specific CNN ArchitecturesKujur et al2., Venkatachalam et al5., Padhi et al34., Kannan et al30.Custom CNN/EfficientNet/compound scalingHigh accuracy for specific cropsLimited cross-crop generalisationNeed for disease-region–centric learningTransfer Learning–Based DetectionSimhadri et al14., Isaac Ritharson et al22., Thakur et al27.Pretrained CNNs fine-tuned on leaf datasetsReduced training time, good performance on small datasetsOver-reliance on whole-image featuresInadequate detection of early or localised lesionsHybrid CNN–RNN ModelsPrommakhot et al1., Ledbin Vini & Rathika3, Yadav et al9.CNN + LSTM/BiLSTM/GRUCaptures contextual and temporal patternsNo explicit lesion localisationMissing ROI-aware temporal modellingROI/Patch/Segmentation-Based MethodsBouacida et al1628.,, Polly & Devi31, Wang & Liu15Patch-based learning, semantic segmentation, object detectionImproved lesion localisation and interpretabilitySpatial-only analysisAbsence of disease progression modellingExplainable AI for Plant Disease ModelsVenkatachalam et al5., Yamaguchi et al6., Upadhyay et al4.Grad-CAM, attention mapsImproved trust and transparencyPost-hoc explainability onlyExplainability not embedded into model designMultimodal & IoT-Driven SystemsLee & Yun11, Pero et al12., Rashid et al29.Image + environmental/IoT sensor fusionEnables early risk predictionIncreased system complexityLack of unified spatial–temporal disease learningReal-Time & Edge-Compatible ModelsHerterich et al8., Joseph et al24., Demilie19Lightweight CNNs, NPU optimizationLow-latency, deployable systemsAccuracy trade-offsNeed for efficient ROI-aware hybrid architectures Summary of Related Work on Deep Learning–Based Leaf Disease Detection. Preliminaries In this section, we first provide the basic concepts and notations used in our proposed LDDHybridNet. The fault of (2) is that it contains no introduction to the ROI-based image representation and the sequence-aware feature modelling, which are fundamental to our method in Sect. 4. ROI-Based leaf image representation Disease detection from field images can be confounded by background clutter, varying lighting conditions, and large areas of noninfected leaves that dominate the visual cues. To alleviate these effects, the Region of Interest (ROI)–based representation has already been widely employed to restrict feature learning to the disease-susceptible region. Let the input RGB leaf image be as Eq. (1). ROI extraction aims to isolate a compact sub-region that contains the symptomatic leaf area while suppressing irrelevant background information. Formally, ROI extraction can be expressed as a spatial masking and cropping operation as in Eq. (2). where denotes a binary mask identifying leaf tissue and represents geometric normalization such as cropping, resizing, and alignment. The standardisation of ROI representation also enables subsequent feature-extraction stages to capture detailed lesion features, thereby enhancing robustness and early-stage disease detection. Sequential modeling of contextual feature dependencies Plant disease appears to exhibit progressive spatial patterns, such as lesion growth, texture changes, and severity differences, yet most existing plant disease datasets consist solely of still images. For ordered feature representations, such contextual interrelations can be learned using sequential modelling approaches, especially recurrent neural networks. Let denote an ordered set of features embeddings derived from spatial partitions or structured representations of an ROI. A Long Short-Term Memory (LSTM) network models’ dependencies across this ordered sequence by maintaining hidden and cell states that selectively preserve salient contextual information. The LSTM update can be expressed as in Eq. (3). where represents the contextual embedding at step . This formulation allows learning progression-aware representations without knowledge of the true time-based supervision, thereby making it applicable to disease datasets that are not based on serial imaging. This step of contextual modelling complements the spatial feature extraction, which models relationships among lesion patterns that are not explicitly captured by these aggregated feature maps. Materials and methods Here we outline the datasets employed, the preprocessing techniques followed, and the architectural abstractions of the LDDHybridNet framework itself. It elaborates on the extraction of return on investment, the design of a hybrid CNN–LSTM model, and training configuration and experimentation protocols. This is motivated by our objective to make the proposed method reproducible and the observation that the proposed method consistently (over multiple experiments) yields similar results over lab and field data collection. Datasets and acquisition To validate that the LDDHybridNet framework is robust and practical, we trained and evaluated it on a small-scale, but diverse, representative subset of publicly available datasets. The choice of these datasets helps train a model suitable for a home lab environment. Still, it generalises better at the field level across challenging environmental conditions such as background clutter, occlusion, and lighting. The main dataset used for training is the PlantVillage dataset43, a comprehensive, high-quality collection of 54,738 images of healthy and diseased leaves from 14 crop species and 38 disease classes. This allows these images to be produced in a controlled lab environment, enabling pretraining and ablation studies on a base model. The images are of fixed size and taken under the same lighting conditions, making it a noiseless platform for learning spatial disease pattern recognition (Gupta). To test generalisation in the field, it uses the FieldPlant dataset44. It includes approximately 5,170 real images for research, taken in a real farming field under normal ambient light, with shadows, background noise, and leaf occlusions. It contains expert-annotated photos of leaves for each disease type (expected acropImage); here, crop examples are Tomato, Cotton, and Grape. Such a way to challenge performance ensures the robustness of the proposed framework in less-than-ideal conditions once deployed in the field. But our model benefits from the joint use of these two datasets: it is trained on clean signals. It does not struggle with slight realistic distortions during the testing phase, as they are what would be encountered in field-deployment scenarios, thereby aiding generalisation. System overview We propose LDDHybridNet, a region-based deep learning framework to detect leaf diseases in precision agriculture in an accurate and interpretable manner while tackling the challenges of (i) background clutter; (ii) subtle or invisible lesions; and (iii) early recognition of leaf diseases. It relies on spatial learning guided by ROI and contextual feature modelling to represent disease when recognised from static field images robustly. LDDHybridNet follows an end-to-end pipeline having five interrelated stages: (i) image input through image acquisition, (ii) image preprocessing and ROI extraction, (iii) deep spatial feature representation via a CNN backbone, (iv) contextual dependency modeling with an LSTM module and ultimately, (v) supervised disease classification with post-hoc explainability. We collect raw images from state-of-the-art benchmark leaf disease datasets that contain different disease classes, lighting conditions, and occlusion situations. These images undergo normalisation and generic noise suppression to stabilise the following processing steps. ROI extraction is an important preprocessing step, during which disease-relevant regions are separated. The background complexity is then suppressed using edge-aware masking techniques whilst preserving the structurally significant symptomatic regions of the leaf. The ROI-masked output image is then passed into a CNN encoding network that converts spatial patterns from small areas of the image into a low-dimensional feature vector that describes important textural, colour, or shape information of the lesion. Figure 1 Each CNN embedding generates a vector for each local patch, which itself encodes spatial features. To account for contextual dependencies across those spatial features, the CNN embeddings are sequenced to form an ordered sequence of spatial feature vectors, indicating a row-wise and column-wise progression pattern across the corresponding ROI. This ordering isn’t a feature of the temporal observations, but rather encodes spatial dependencies/structure. This ordered representation is processed by an LSTM network, which propagates contextual information across spatial regions, allowing the model to learn progression-aware feature relationships characteristic of both disease severity and distribution. End-to-End System Architecture of the LDDHybridNet Framework for Early and Accurate Leaf Disease Detection Using ROI-Enhanced CNN-LSTM Modelling and Explainable Predictions. Finally, the aggregated CNN–LSTM representation is fed to a dense classification head that predicts the associated leaf disease class. We employ a Grad-CAM-based explainability module after the inference step to better highlight the ROI regions that most influence the model’s output, enabling greater transparency and expert interpretation. The significance of Grad-CAM, however, is that it functions as a post-hoc visualisation method and is not involved in either feature learning or decision-making. The overall end-to-end architecture of the LDDHybridNet framework, demonstrating the major components and how data flows among them, is depicted in Fig. 1. This organised structure provides accurate detection of leaf diseases, interpretable results, and can be used for practical purposes in the field of precision agriculture. In Table 2, we summarise the notations used for inputs (i.e., ROI), CNN-LSTM embeddings, and explainability variables within the LDDHybridNet framework. Table 2Notations Used in the LDDHybridNet Framework.NotationDescription Input RGB leaf image Region of Interest extracted from the input image Feature extractor function of LDDNet CNN backbone Output spatial feature map from CNN ROI-focused feature map after attention masking Time-series or sequential feature embedding for LSTM Hidden state vector at time step in LSTM Cell state vector at time step in LSTM Weight matrices for forget, input, and output gates in LSTM Grad-CAM activation map for visual explanation Importance weight of feature map for class in Grad-CAM feature map from last convolutional layer Class score for class Final predicted label Model prediction probabilities (softmax output) Notations Used in the LDDHybridNet Framework. Data preprocessing and augmentation A robust preprocessing and augmentation pipeline is applied to both the PlantVillage43 and FieldPlant44 datasets to maintain consistency and further improve the model’s generalisation capability. Because all convolutional layers in the LDDHybridNet model accept fixed-sized input leaf images of size 224 × 224 × 3 pixels, all input leaf images are resized to this size. Min-max normalization is applied to pixel values to bring them into a range of [0, 1] as expressed in Eq. (4). is the original pixel intensity and is the normalized pixel intensity. This oaring maintains shorter convergence time while nullies light variations amongst datasets. We perform median filtering with 3 × 3 kernel to de-noised images especially when it is real fields data. The value of the filtered pixel can be calculated as in Eq. (5). where is the neighborhood of pixel . This provides a good balance between disease relevant edge preservation and suppression of the salt-and-pepper noise. Contrast also increases with the local contrast using Contrast Limited Adaptive Histogram Equalization (CLAHE) which divides the image into small tiles and processes histogram equalization for each tile individually. The contrast limited adaptive histogram equalization (CLAHE) enhances feature visibility in low lit areas/low contrast areas which aids in the detection of early stage symptoms by the model. Various transformations such as random horizontal and vertical flipping, ± 15° rotation, up to 10% zoom shifts, and minor translations are performed to augment the training set and mimic the variability as seen in the real world. The augmentations take into account orientation differences, scale distortions and leaf occlusions as in real outdoors scenarios. We also introduce CLAHE-enhanced version of the images to make the model more robust under different lighting conditions. One of the main problems in plant disease datasets is the class imbalance especially for rare or first stages symptoms. It combats this by synthetically generating samples of underrepresented classes with a Generative Adversarial Network (GAN). Given random noise vectors z, the generator network generates synthetic disease images as in Eq. (6). This creates new synthetic data that enhances the minority classes and mitigates class imbalance and overfitting to majority classes. The dataset is divided into training (80%), validation (10%), and testing (10%) sets following preprocessing and augmentation. This augmentation was only applied to the training set to ensure evaluation remained unbiased. Preprocessing methods are equally implemented on each subsets to homogenate input form. ROI extraction module As part of the LDDHybridNet pipeline, a Region of Interest (ROI) extraction mechanism is introduced as a core preprocessing module to explicitly focus learning on disease-relevant leaf regions. This module is designed to suppress background clutter, localize symptomatic areas, and preserve fine-grained texture and color cues that are critical for early-stage disease detection. Given an input leaf image . leaf tissue is first segmented from the background using a U-Net–based segmentation model. The U-Net is fine-tuned on leaf mask annotations derived from the PlantVillage dataset and produces a binary segmentation mask . The segmented leaf image is obtained via pixel-wise masking: where denotes element-wise multiplication. This operation retains only the leaf tissue while setting background pixels to zero. Figure 2 The input leaf image is first segmented using a U-Net model to obtain a binary leaf mask, from which a tight bounding box is deterministically derived to localise the diseased region. The extracted ROI is then resized to a fixed resolution (128 × 128), orientation-aligned via affine transformation, contrast-enhanced using CLAHE, and augmented before being forwarded to the CNN backbone for feature extraction. ROI Extraction and Preprocessing Workflow for Leaf Disease Localization. Following segmentation, a tight bounding box is computed directly from the binary mask to enclose the diseased leaf region. Unlike earlier descriptions, YOLO is not used as an independent detection stage; rinstead, bounding box extraction is performed deterministically from theU-Net mask geometry. The resulting bounding box defines the ROI, which is then resized and padded to a fixed spatial resolution of to ensure architectural consistency throughout the network. All subsequent processing stages operate on this standardised ROI size. To further reduce intra-class variability caused by leaf orientation and capture conditions in field images, an affine alignment step is applied to the extracted ROI. The principal orientation of the leaf is estimated using second-order image moments, and the ROI is rotated using a transformation matrix as in Eq. (5). where denotes the cropped region defined by the bounding box . This alignment mitigates the effects of leaf slant, rotation, and pose variation, thereby stabilizing feature learning. Then, Contrast Limited Adaptive Histogram Equalisation (CLAHE) is applied to the grayscale version of the ROI to improve local contrast and enhance subtle lesion patterns, which would otherwise remain imperceptible in early infection stages. All mild transformations, such as rotations, flips, and intensity jittering, are applied during training only to the region of interest (ROI) and preserve the semantic consistency of lesion areas. At the end, the processed ROI is then passed to the CNN backbone of LDDHybridNet to extract its spatial features. The accompanying ROI-guided features are subsequently subjected to contextual modelling via the LSTM module, enabling disease classification in a progression-aware, symptom-centric manner. The proposed method ensures a stable and clear ROI extraction pipeline that minimises false activations from non-cancerous regions and is robust to different imaging conditions. LDDHybridNet architecture To overcome these challenges, we propose LDDHybridNet, a sequential image-regions-of-interest (ROI)-aware convolutional neural network (CNN)–long short-term memory (LSTM) architecture for early and accurate leaf disease detection using static images. Early use of the word ‘parallel’ to describe the proposed model is misleading; instead, it follows a sequential processing pipeline, in which the spatial feature extraction, contextual dependency modelling, and attention-based fusion components occur in order, with attention-based fusion at the end of the process. The architectural flow is consistently defined as ROI extraction → CNN feature encoding → attention-guided sequential modeling → classification, which is illustrated in Fig. 3. Model Architecture of LDDHybridNet with Attention and Temporal Modules. The architecture begins with a Region of Interest (ROI) image obtained from the preprocessing stage, which isolates lesion-dominant regions and suppresses background interference. Each ROI image is passed through a compact CNN backbone inspired by the LDDNet encoder. This backbone is intentionally lightweight to preserve fine-grained disease cues while maintaining computational efficiency. The CNN produces a spatial feature vector . which encodes local texture, color irregularities, and lesion morphology. While the repositories used in this study (e.g., PlantVillage and FieldPlant) are based on static images rather than real-time series data, modelling the contextual dependency remains helpful in capturing the variance in disease severity and diffusion within the lesion area. In this way, we use a synthetic-sequential representation by treating several feature embeddings from ordered ROIs (or spatially partitioned feature vectors) as a sequence. Rather than assuming real temporal evolution, this formulation models contextual progression cues encoded in the arrangement of spatial features. Formally, the ordered feature embeddings . are provided as input to a Long Short-Term Memory (LSTM) module, which captures dependencies across the ordered sequence. At each step , the LSTM computes a contextual embedding: where denotes the hidden state from the previous step. This mechanism enables the model to learn progression-aware representations without relying on longitudinal image acquisition. To emphasize feature embeddings that exhibit stronger disease characteristics, an attention-based fusion mechanism is applied over the LSTM outputs. Temporal (contextual) attention weights are computed as in Eq. (7). assigning higher importance to embeddings that are more discriminative of disease presence or severity. The final spatio-contextual representation is obtained as in Eq. (8). which summarizes lesion characteristics and contextual dependencies across the ordered feature sequence. This fused representation is passed through fully connected layers followed by a Softmax classifier as in in Eq. (9). where and are the output layer parameters. Overall, LDDHybridNet is a single-stream, context-aware architecture that combines ROI-guided spatial learning with sequence-based contextual modelling under the constraints of static-image datasets. By avoiding claims of temporal supervision and clearly defining how sequential modelling is synthesised, the proposed design ensures architectural transparency and methodological integrity. Its modular structure supports end-to-end training and can be extended to actual longitudinal crop monitoring scenarios when genuine time-series data becomes available. Loss function and training strategy The proposed LDDHybridNet framework is trained end-to-end in a supervised manner to minimise classification error across disease categories while ensuring stable convergence under class-imbalanced conditions. For multi-class disease classification, the categorical cross-entropy loss is employed, which is standard and effective for probabilistic Softmax-based outputs. The loss function is defined as: where denotes the number of disease classes, is the one-hot encoded ground truth label, and is the predicted probability for class . To address class imbalance, which is common in real-world leaf disease datasets, a class-weighted loss formulation is adopted. Each class is assigned a weight inversely proportional to its frequency : ensuring that minority disease classes contribute proportionally more to the optimization objective. The resulting weighted loss over a mini-batch of samples is computed as: Model optimization is performed using the Adam optimizer with an initial learning rate of . A learning rate scheduler is applied to reduce the learning rate by a factor of 0.1 when validation loss fails to improve for five consecutive epochs. Training is conducted with a mini-batch size of 32 for a maximum of 100 epochs, with early stopping after 10 epochs based on validation performance. Batch normalisation and dropout regularisation (dropout rate p = 0.3) are used to prevent overfitting. Input ROIs are resized to 128 × 128, normalised to zero mean and unit variance, and subjected to horizontal flipping, small rotations, and brightness adjustments to improve generalisation. TensorFlow 2.0 is used to implement the full training pipeline, which runs on NVIDIA RTX-series GPUs (16 GB VRAM) with 32 GB system memory. Our strategy for training not only provides numerical stability but also demonstrates robustness to class imbalance and consistent convergence across both lab- and field-based datasets. Design rationale and architectural justification LDDHybridNet is motivated by a problem-driven architectural philosophy rather than a heuristic stack of popular deep learning modules. Three core challenges arise in early leaf disease detection for precision agriculture: (i) subtle leaf lesion patterns typically lie within complex visual backgrounds, (ii) there is an innate sequential nature of disease progression, yet temporal annotations are rarely explicit, and (iii) agronomic trust and adoption require transparent decision-making. We introduce components within the proposed framework, each of which targets one of these challenges without overlap and with a clear functional motivation. Initially, we use ROI-aware preprocessing to spatially limit feature extraction to the leaf regions relevant to the disease. Previous works have also shown that symptoms can be blurred in early images due to background noise and non-informative leaf areas. By serving as a locality before isolating lesion-dominant regions, the model reduces computational overhead and can learn from fine-grained pathological cues, thereby improving sensitivity to early manifestations of the disease. Secondly, since convolutional feature extraction enables the effective capture of localised texture, colour variation and edge formation regarding leaf diseases, it is used as the primary spatial encoding strategy. Instead of having different levels of attention or multi-scale enhancement modules stacked on top of each other, it has an intentionally lean backbone that retains discriminative power without introducing a computationally expensive architecture. This selection provides sufficiently robust spatial representations while still being efficient enough for precision agriculture workflows. Third, an additional Long Short-Term Memory (LSTM) layer is incorporated to capture the implicit dynamics of disease progression. While specific leaf images likely do not directly encode time-series data, the spatial patterns associated with disease time-series evolution (e.g., lesion spread, morphological change, and severity gradients) can be inferred from spatial images across regions of interest. The LSTM allows the network to learn relationships between the extracted spatial features, thereby obtaining progression-aware representations without requiring longitudinal image sequences. Such a design provides a computationally efficient approach to temporal modelling while maintaining the capacity for context reasoning based on the transformer. Several of the issues addressed in the framework are not entirely new to research, but, importantly, they treat explainability as a separate design objective rather than a trade-off-affecting component. Only for the sake of visualising the rationale behind the trained model’s decisions, we incorporate Gradient-weighted Class Activation Mapping (Grad-CAM), ensuring that biologically meaningful lesion regions drive the predictions. Grad-CAM does not affect feature learning or classification predictions; instead, it provides a validation process to confirm that our model remains interpretable visually and agronomically. As such, the derived architecture is a lightweight but discrete combination of ROI-driven visual-spatial learning and multi-level temporal contextual modelling, not during training explanation. Each module has its specific function, and we don’t use any components for redundancy or trend-based designs. In contrast, existing approaches stack modules without a clear rationale, whereas LDDHybridNet integrates modules in a principled manner through its design. In addition, following ablative observations corroborate that removing ROI guidance or substituting LSTM with a spatial pooling layer consistently results in performance deterioration, reiterating the indispensability of these components in our architecture. Algorithmic representation The subsequent sections rigorously describe the complete algorithmic flow of the LDDHybridNet framework. It summarises the key steps of ROI-based preprocessing, CNN feature extraction, attention-based enhancement, and the LSTM temporal model. The algorithm specifies each step in the detection pipeline to help implement and reproduce the hybrid deep learning strategy. Algorithm 1: LDDHybridNet – ROI-Aware CNN–LSTM Based Leaf Disease Detection Procedure Algorithm 1 summarizes the complete training and inference workflow of the proposed LDDHybridNet framework. The input leaf image is first preprocessed and refined through ROI extraction to suppress background noise and isolate disease-relevant regions. A CNN backbone then encodes fine-grained spatial features from the ROI. Since the available datasets consist of static images, sequential modeling is achieved by constructing an ordered set of spatial features embeddings derived from the CNN feature maps rather than true temporal observations. These ordered embeddings are processed by an LSTM to capture contextual dependencies that reflect spatial progression patterns of lesions. The aggregated LSTM representation is finally classified using a Softmax layer. Grad-CAM is employed only during post-hoc analysis to visualize lesion-sensitive regions and does not participate in model training. Evaluation Metrics and Experimental Setup To effectively measure the performance of the proposed LDDHybridNet framework and the leaf disease classification results, several standard classification metrics are used. These metrics capture the overall accuracy of prediction as well as how well a class can be predicted, which is important for imbalanced agricultural datasets. Accuracy, the overall correctness, is defined as in Eq. (16). where True Positives, True Negatives, False Positives, and False Negatives are denoted by TP, TN, FP, and FN, respectively. Precision (P) measures the percentage of positive predictions that are correct as in Eq. (17). Recall tells us how well the model identifies real positives as in Eq. (18). F1-score gives the harmonic mean of precision and recall, balancing between false positives and false negatives as in Eq. (19). AUC–Area Under the Curve assesses how well the model can differentiate between the different classes, especially for imbalanced dataset. The greater the AUC, the better the discriminative ability. In addition, to interpret class-wise prediction behaviour via a Confusion Matrix, the literature also provides interpretations of the actual vs. predicted distributions across different disease categories. The stratified 5-fold cross-validation protocol used to train and evaluate the LDDHybridNet model provides a reliable and generalizable estimate of performance. We divide the baseline dataset of 100 into 70% for training, 15% for validation, and 15% for testing. During training, early stopping and learning rate scheduling work hand in hand to help the network converge to the best point. Augmented samples are only applied to the training set to avoid leakage. All experiments are performed in a TensorFlow 2.0 environment on a computer with an NVIDIA RTX 3080 GPU (10GB VRAM), an Intel i7 CPU, and 32GB RAM. We report the mean ± standard deviation for each metric across 3 runs with different random seeds to improve statistical robustness. 10 Explainability and Visual Validation The LDDHybridNet framework, therefore, combines explainability techniques that are crucial for building agronomically reliable trust in model predictions. While conducting experiments, we are using the Grad-CAM method to visualise the regions in leaf images that the model focuses on for classification. Grad-CAM produces coarse localisation heatmaps from the gradients of the target class flowing into the final convolutional layer. These gradients provide us with the most important pixels, helping a lot in the final prediction, so it marks the region of interest (ROI) where the disease manifests. We compute the Grad-CAM activation map for a class c as in Eq. (20). with denotes the feature map of the last layer of the convolutional part of the neural network, and is the importance weight obtained by averaging over all positions of the gradients of the score for class c with respect to feature map as in Eq. (21). They are superimposed on the original images to confirm that the model focuses on biologically relevant disease-infected regions rather than regions considered background noise or irrelevant. This enables agreement between the human and the model and could also enable expert verification in precision agriculture. Saliency maps and class-wise confusion matrices are combined for visual validation, aiding further understanding of the model’s behaviour under ambiguous or mixed symptom cases in agronomy. Experimental results The experimental results evaluate the effectiveness, robustness, and early detection capability of the proposed LDDHybridNet framework across both controlled and real-world agricultural conditions. Comprehensive quantitative metrics, class-wise analyses, ablation studies, and explainability assessments are presented to demonstrate how ROI-guided spatial learning and temporal modelling significantly enhance disease recognition performance compared with traditional CNN-based and transformer-based baselines. Experimental setup and implementation details We implemented all our experiments using TensorFlow 2.x on a workstation with an NVIDIA RTX 3080 GPU (10 GB VRAM), an Intel Core i7 processor, and 32 GB of system memory. The proposed LDDHybridNet framework is trained in a supervised end-to-end manner. All experiments use a 5-fold CV protocol to provide a robust and unbiased estimate of performance (w.r.t. PlantVillage ~ 54k leaf images, FieldPlant: 5,170 images taken in the field under unconstrained conditions). Mutable folds (e.g., for each dataset, samples are stratified by class and partitioned into five exclusive folds). For all cross-validation folds (5-fold CV), 3 folds (60%) are used to train the networks, 1-fold (20%) for validation, and 1-fold (20%) to compute the final results, which are reported as the average across all folds. All preprocessing and augmentation steps are applied after fold construction to avoid data leakage. Only the training folds are augmented, while validation and test folds remain unchanged. The datasets do not provide explicit plant- or scene-level identifiers; therefore, strict plant-wise or scene-wise separation cannot be enforced, which is acknowledged as a limitation of the evaluation protocol. Region-of-Interest (ROI) images extracted during preprocessing are resized to a fixed resolution of 128 × 128 × 3 and normalised to the range . The model is optimized using the Adam optimizer with an initial learning rate of , , , and a weight decay of . A mini-batch size of 32 is used, and training is performed for a maximum of 100 epochs, with early stopping (patience = 10 epochs) based on validation accuracy. If the validation loss does not improve for five consecutive epochs, a learning rate scheduler reduces the learning rate by a factor of 0.1. To enhance generalisation, standard data augmentation techniques—including horizontal and vertical flipping, rotations within ± 15°, zooming, and brightness jitter—are applied only to the training folds. In addition, GAN-based data augmentation is employed to mitigate class imbalance. Synthetic samples are generated exclusively from real training images of minority classes within each fold. They are never introduced into the validation or test folds, ensuring no performance inflation from synthetic data leakage. For each minority class, synthetic images are generated to approximately balance class distributions within the training fold. Regularisation is further enforced via dropout (p = 0.4) and batch normalisation layers throughout the network. To account for training variability, all reported performance metrics are presented as mean ± standard deviation over three independent runs with different random seeds. As seen in Fig. 4, the number of images per PlantVillage disease class is highly unbalanced. There is a mild imbalance: some classes have more than 3,000 images, while others have fewer than 300. Such variation also means we need robust training strategies, such as carefully designed loss functions and data augmentations, to ensure a stable, unbiased model. Class Distribution of PlantVillage Dataset Across 15 Disease Categories. Focusing on the PlantVillage dataset, Fig. 5 shows a range of representative images of the healthy and diseased leaf categories for early blight, late blight, bacterial spot, mosaic virus, leaf mould, and spider mite infestation. The samples exhibit diversity in colour, texture, and lesion morphology, demonstrating the dataset’s diversity and suitability for training a leaf disease classification model that generalises, such as our LDDHybridNet. Representative Sample Images from the PlantVillage Leaf Disease Dataset. As shown in Fig. 6, the FieldPlant contains 60,648 images, with a wide variation in sample sizes across both crop types and disease categories. The large number of classes includes TomatoBrown Spots, Cassava Mosaic, and Corn Leaf Blight, while some classes have very few samples. This degree of imbalance mimics field conditions in the real world while also highlighting the need for modelling approaches that are flexible enough to handle uneven, often noisy, and complex agricultural data. Class Distribution of the FieldPlant Dataset Across Diverse Crop Diseases and Conditions. A detailed pie chart illustrating the proportional distribution of the FieldPlant dataset is shown in Fig. 7. Tomato Brown Spots, Cassava Mosaic, and Corn Leaf Blight comprise the majority of the dataset, accounting for more than half of all samples. This imbalance is further reflected in several classes that contribute less than 1%. This distribution mimics real-field conditions and emphasises the importance of powerful models that can learn from imbalanced and heterogeneous agricultural data. Proportional Class Distribution of the FieldPlant Dataset Across Multiple Crop Diseases. Baseline models and comparative framework To rigorously evaluate the effectiveness of the proposed LDDHybridNet framework, it is compared with a variety of leading state-of-the-art models for leaf disease detection and image classification, including convolutional, transformer-based, hybrid, and ROI-aware models. We choose these baselines such that they cover all the cases where the contribution of (i) ROI-aware preprocessing, (ii) multi-scale and attention-based spatial encoding, and (iii) temporal modelling can be isolated. The first is the standard CNN backbones, where each network is trained end-to-end on full images without performing ROI extraction, namely VGG16, ResNet50, DenseNet121, MobileNetV3-Small, and EfficientNet-B0. They provide a solid foundation for adopting deep feature-extraction-based approaches across a range of computational budgets and are well documented in the literature for plant disease classification. They provide examples of the distances these well-designed, lightweight CNNs can achieve without explicit modelling of either time or ROIs. Secondly, to assess the advantages of global context modelling using transformers, we introduce a Vision Transformer (ViT-B/16) baseline trained on 224 × 224 leave images. Such a model learns long-range dependencies and patch-level relations, enabling comparisons between pure self-attention architectures and the CNN–LSTM hybrid design employed in the LDDHybridNet approach. A CNN–LSTM baseline, which utilises a standard ResNet-based encoder followed by a unidirectional LSTM, is also implemented to evaluate the advantages of leveraging bidirectional temporal modelling and attention-based fusion. Finally, we compare with the most similar frameworks to our setting: ROI-aware and hybrid frameworks. These were the previously proposed LeafDiseaseNet (ROI-unaware CNN with SE blocks), as well as the HybridLeafDiseaseNet (ROI-based hybrid CNN–ASPP–CSAM–LSTM backbone). We also evaluate an ROI + CNN-only variant (U-Net + YOLO + CNN, without LSTM or temporal attention) to separate the role of temporal modelling. Collectively, these baselines provide an extensive comparison framework, establishing that the advantages in accuracy, robustness, and early-stage disease detection provided by the explicit incorporation of ROI-aware spatial encoding, multi-scale attention, and temporal sequence learning with LDDHybridNet are consistent within controlled laboratory and field environments. Quantitative performance evaluation For the quantitative evaluation, we employed standard metrics, namely Accuracy, Precision, Recall, F1-Score, and AUC, on LDDHybridNet using the PlantVillage and FieldPlant datasets. The results in Table 3 showed that the model achieved competitively high scores not only on clean generalisation but also on real-world generalisation, demonstrating the effectiveness of ROI extraction, multi-scale attention, and temporal modelling. LDDHybridNet has attained 98.91% accuracy on the PlantVillage dataset and generalises well, achieving 95.37% accuracy on the FieldPlant dataset despite various background noises, occlusions, and illumination differences. Table 3Quantitative Performance of LDDHybridNet on PlantVillage and FieldPlant Datasets.DatasetAccuracy (%)Precision (Macro)Recall (Macro)F1-Score (Macro)AUC (Macro) PlantVillage 98.91 98.7298.6498.670.996 FieldPlant 95.37 94.8294.3594.510.982 Quantitative Performance of LDDHybridNet on PlantVillage and FieldPlant Datasets. Macro-averaged values of Precision, Recall, F1-Score, and AUC were calculated to compare performance across all disease categories in out-of-balance detection. Also, confusion matrices show higher true-positive rates and lower misclassification rates in visually related disease pairs. Our comparative bar charts reveal significant performance improvements over the CNN-only, transformer-only, and hybrid baselines, highlighting the benefits of integrating ROI-aware spatial encoding with temporal modelling. In summary, the proposed LDDHybridNet is shown to be a highly reliable model for application in either controlled environments or precision agriculture fields. The PlantVillage dataset is a curated, laboratory-style benchmark with limited background variability; consequently, very high classification accuracies approaching a performance ceiling have been reported across multiple deep learning studies. To avoid inflated performance estimates, all dataset splits were performed before data augmentation, and augmentation was applied exclusively to the training set, while the validation and test sets remained unaugmented. In addition, preprocessing operations were strictly label-agnostic and did not use disease annotations beyond class labels. While LDDHybridNet achieves high accuracy on PlantVillage, this result is interpreted in light of the dataset’s ceiling effect, and greater emphasis is placed on field dataset results to demonstrate robustness under real-world conditions. We summarise the overall performance of LDDHybridNet on the PlantVillage and FieldPlant datasets in terms of Accuracy, Precision, Recall, F1-Score, and AUC in Fig. 8. The five-evaluation metrics are shown on the X-axis, and their corresponding values are displayed on the Y-axis. It consistently scores well in both synthetic and real domains, demonstrating performance benefits in undistorted images and considerable robustness to field distortions. This figure visualises the discriminative ability of LDDHybridNet for various agricultural backgrounds. Comparative Analysis of Overall Performance Metrics Across Datasets. The F1-Score over all classes for each disease category of PlantVillage and FieldPlant is illustrated in Fig. 9. In this plot, the actual names of diseases are written on the X-axis, and their F1 Scores are labelled on the Y-axis. The performance of LDDHybridNet on field diseases is consistent with that on controlled data, but with lower scores due to occlusions and illumination. The model shows strong interclass discriminability between tomato, corn, and grape leaves, indicating its effectiveness for multiclass identification of different fungi. Class-Wise F1-Score Comparison Between PlantVillage and FieldPlant. The confusion matrix for LDDHybridNet on the PlantVillage dataset based on the actual disease names is shown in Fig. 10. The X-axis and Y-axis are predicted classes and accurate classes, respectively. Thus, colour intensity and numerical values inside each cell provide a correct measure of classification accuracy. However, robust diagonal dominance indicates very high predictive reliability under controlled conditions. For the five major disease categories, ROI-guided CNN–LSTM modelling achieves high performance (AUC > 0.982) and demonstrates strong discriminative capability (minimal off-diagonal values). Confusion Matrix of LDDHybridNet on the PlantVillage Dataset. We report the confusion matrix of LDDHybridNet on the FieldPlant dataset in Fig. 11, with the numbers written in all cells. The predicted classes appear on the X-axis and the true labels on the Y-axis. To show that diagonal values dominate despite noise in the other channel (occlusions, lighting, similar reflection planes, etc.), diagonal values show good field-level generalisation. These misclassifications are clustered in diseases with overlapping symptoms, but the overall accuracy is very high. This upgraded visualisation effectively illustrates the model’s robustness under realistic agricultural scenarios. Confusion Matrix of LDDHybridNet on the FieldPlant Dataset. Figure 12 LDDHybridNet vs. common baseline accuracy performance. In this section, we compare LDDHybridNet with common baseline architectures, including CNNs, transformers, and other hybrid models. Models are listed along the X-axis, and classification accuracy along the Y-axis—the value labels at the top of each bar indicate the rationale for the performance comparisons. LDDHybridNet achieves the highest performance (98.91%) and outperforms classical deep networks and transformer-based methods by large margins. The results illustrate the ability of ROI extraction, multi-scale attention, and temporal modelling to detect leaf disease quickly and accurately. Accuracy Comparison of LDDHybridNet Against Baseline Models. Class-wise classification performance The performance of LDDHybridNet was analysed for discrimination between different classes of plant diseases. Accordingly, precision, recall, and F1 − score were calculated per class for five example diseases from the PlantVillage and FieldPlant datasets. These results also provide insight into the strengths and weaknesses of the model: distinguishing between closely related infections with similar appearances vs. fine distinctions in lesion subtleties. These results indicate consistent high performance across all classes and reliable generalisation in the field. Since the overall reported numbers in the previous sections correspond to macro—averaged values, they are summed and highlighted in the “Aggregated” row. The Class-wise LDDHybridNet Precision, recall, and F1-score on five disease categories are shown in Table 4. The high values of these metrics, which are stable across the classes, imply strong discriminative power and a well-balanced classification. The consistency appears strong overall, as indicated by the macro-averaged row: the model generalises well to changing lesion presentations, maintaining high prediction accuracy. Table 4Class-Wise Precision, Recall, and F1-Score for LDDHybridNet (PlantVillage + FieldPlant Combined).Disease ClassPrecision (%)Recall (%)F1-Score (%)Tomato Early Blight (TEB)98.698.498.5Tomato Late Blight (TLB)99.198.798.9Corn Common Rust (CCR)98.197.998.0Corn Leaf Blight (CLB)98.898.598.6Grape Black Rot (GBR)98.398.098.1Total (Macro Average)98.5898.3098.42 Class-Wise Precision, Recall, and F1-Score for LDDHybridNet (PlantVillage + FieldPlant Combined). Class-Wise Precision, Recall, and F1-Score Bar Plot. Basic information regarding Tomato, Corn, and Grape. The class-wise Precision, Recall, and F1-Score for our five main categories of diseases common on tomato, corn, and grape plants are shown in Fig. 13. The disease classes are indexed along the X-axis, and the metric values are plotted along the Y-axis. The high, balanced performance across all metrics throughout the epochs is shown in the visualisation, reinforcing that LDDHybridNet detects both small and large lesions. A slight variation across classes indicates a high generalisation quality and shows that ROI-guided spatial encoding and temporal modelling are effective in differentiating visually similar diseases. Class-Wise Performance Heatmap. As illustrated in Fig. 14, heatmaps of class-wise Precision, Recall, and F1-Score provide an easily interpretable overview of the model’s discriminative capability across various disease categories. Rows are disease classes, while columns are evaluation metrics. The colour intensities and overplotted numbers indicate uniformly strong performance, and any slight variations from this pattern are to be expected given the common overlap of symptoms. The recorded heatmap shows that LDDHybridNet is a reliable prediction tool, as the confidence and robustness of the predictions do not change across different disease types, and across several learned conditions (in vitro, in vivo, and in clinic). ROI Extraction effectiveness analysis Next, to measure the effect of progressively refined ROI extraction on leaf disease recognition, three configurations were tested: (i) Without ROI, which processes the complete leaf image by the network; (ii) U-Net Only, where we artificially imposed the foreground segment mask to remove most background noise; (iii) U-Net + YOLO, the proposed two-stage ROI extraction pipeline in LDDHybridNet. The comparison was made under the same training conditions for all configurations. Without an ROI setup, the accuracy is lower due to irrelevant background textures, shadows, soil patterns, and partly obscured leaf parts. U-Net segmentation is applied to enhance the results by focusing only on the leaf region; however, because lesion boundaries are finely defined, regions remain unclear. Finally, the best performance of the full U-Net + YOLO ROI module is achieved by generating tight bounding boxes that focus on lesions, matching their orientation, and highlighting initial-stage spots. It works by sequentially refining spatial attention and providing temporal modelling cues, leading to significant suppression of false activation in asymptomatic regions. We summarise performance improvements in Table 5. The most prominent enhancements were observed with the “U-Net + YOLO” configuration (compared to others) across all metrics, as expected for diseases with typical symptom differences. ROI extraction gradually removed noise, identified disease regions, and improved the accuracy of lesion localisation, as evidenced by qualitative visual comparisons. Table 5Impact of ROI Extraction on LDDHybridNet Performance.ConfigurationAccuracy (%)Precision (%)Recall (%)F1-Score (%)Without ROI94.1293.6593.2193.42U-Net Only (Segmentation)96.8496.4096.1596.22 U-Net + YOLO (Final ROI – Proposed) 98.91 98.72 98.64 98.67 Impact of ROI Extraction on LDDHybridNet Performance. Accuracy for three configurations of ROI extraction in LDDHybridNet (Fig. 15). As the background noise is severe and the lesion focus is not strong, the accuracy obtained is the lowest in the “Without ROI” setting. While U-Net segmentation improves performance by focusing on leaf regions, the best performance is achieved by combining U-Net and YOLO to extract ground truth for the final ROIs. This suggests the need for accurate lesion localisation for consistent classification between disease classes. Accuracy Comparison Across ROI Extraction Configurations. The F1-score after each stage of ROI extraction is shown in Fig. 16. When the entire image is used, distractive backgrounds hinder the model from focusing on the lesioned images it needs to learn. Background elements are removed by U-Net segmentation, leading to an improvement in the F1-score. Combining the U-Net + YOLO pipeline yields the best performance, as it tightly constrains symptomatic regions and helps the classifier discriminate subtle disease patterns. That is a progressive phenomenon that illustrates the urgency of extracting high-level ROI. F1-Score Progression Across ROI Extraction Stages. The precision and Recall of these three ROI extraction approaches are shown in Fig. 17. With its advanced refinement, both these metrics have improved significantly. The baseline without ROI extraction has the lowest lesion sensitivity and the highest confusion. Improved leaf segmentation improves both U-Net segmentation metrics. Combining a full U-Net + YOLO pipeline results in significant benefits, precisely localising symptomatic tissues with high precision and recall. This provides evidence that the central role in enhancing classification robustness lies in ROI extraction. Precision and Recall Comparison for ROI Extraction Methods. Three-stage ROI extraction for a Tomato Bacterial Spot sample is illustrated in Fig. 18 the input image with background noise and different illumination. Results: The U-Net mask is a cleaned area that preserves leaf structure and removes non-leaf regions, and the YOLO overlay shows the exact bounding box that draws down the infected area. The joint pipeline outputs a lesion-centred ROI with high contrast, yielding high accuracy in downstream classification in the LDDHybridNet architecture. ROI Extraction Illustration for Tomato Bacterial Spot Using U-Net Segmentation and YOLO Bounding. ROI extraction workflow for a Tomato Septoria Leaf Spot sample (shown in Fig. 19). The input image has complex lesion patterns with irregular shapes and multi-region infection spread. The U-Net mask shows the diseased regions in a high-fidelity segmentation of the lesion pixels, and the YOLO overlay provides a pure bounding box that contains only the infected leaf area. Such multi-phase representation improves lesion contrast, reduces distracting background texture, and leads to substantially better learning of small but disease-specific visual patterns by the model. ROI Extraction Illustration for Tomato Septoria Leaf Spot Using U-Net Lesion Masking and YOLO Disease Localisation. Natural environment field sample ROI extraction for Corn Common Rust (CCR) (Fig. 20). The input image includes complex backgrounds, varied lighting, and long lesions—u-Net mask segments rust pustules on the leaf, leaving only the infected rust regions. The YOLO overlay will draw a high-precision bounding box around the disease area. This sequential pipeline has the potential to clear the field of unwanted backgrounds, make lesions more perceptible, and facilitate better downstream classification in complex outdoor environments. ROI Extraction Illustration for Corn Common Rust (CCR) Using U-Net Lesion Segmentation and YOLO Field-Level Localisation. In this manner, see Fig. 21, the ROI extraction process in a Tomato Target Spot sample. An average input sample is an image of several circle-like patterns of lesions, with blurred inputimages that cover their boundaries, making them hard for classic classifiers. Lesion ROIs are delineated by a U-Net mask, which not only demarcates the necrotic concentric rings but also produces distinct lesion features. Finally, the infected area is covered with a clean bounding box using the YOLO overlay and revising the ROI. We observed that the optimal combination reduces background noise, maximises lesion visibility, and trains LDDHybridNet to focus on extracting only diagnostic information features. ROI Extraction Illustration for Tomato Target Spot Using U-Net Lesion Segmentation and YOLO ROI Refinement. Flowchart for the extraction of the (ROI) of the Tomato Early Blight sample (Fig. 22). The input image shows irregular necrotic lesions with varying colour intensity and diffuse margins. While the U-Net mask segments highly sensitive clusters of lesions and removes them from surrounding non-infected regions, the YOLO ROI overlay creates a precise bounding box that encapsulates the entire diseased region. Such a combined strategy reduces background noise, increases the lesion’s observability, and helps LDDHybridNet learn relevant features for early blight discrimination. ROI Extraction Illustration for Tomato Early Blight Using U-Net Lesion Segmentation and YOLO Bounding Refinement. Ablation study of LDDHybridNet components An ablation study was performed to quantify the contributions of the main parts of the LDDHybridNet framework. We developed six variants of our model by gradually removing critical modules: (i) SE Block, (ii) Convolutional Block Attention Module (CBAM), (iii) ViT Head, (iv) ASPP multi-scale extractor, (v) Bi-LSTM temporal encoder, and (vi) Temporal Attention Mechanism. All other parameters were kept constant, except execution time, to isolate the impact on their performance. It can be seen that dropping any one of them would result in a statistically significant decrease in accuracy, F1-score, and AUC. However, the most significant drop is observed when performance without temporal modelling (Bi-LSTM and temporal attention) is evaluated, indicating the significance of incorporating the sequential evolution of lesion patterns within the ROI-attention-based input sequence, and that multi-scale representation and long-range feature encoding by the ViT head model and ASPP modules are also crucial for performance. In contrast, SE blocks and CBAM provide noticeable, but more limited, improvements via channel/spatial reweighting and localised attention refinement. The summary of all results is listed in Table 6. Table 6Ablation Results for LDDHybridNet Components.Model VariantAccuracy (%)F1-Score (%)AUC (%)Full LDDHybridNet (Proposed)98.9198.6799.12Without SE Block97.8497.4098.21Without CBAM97.5297.1197.93Without ViT Head96.4896.0296.87Without ASPP95.9195.3396.15Without Bi-LSTM94.7294.1095.26Without Temporal Attention94.3893.8294.97 Ablation Results for LDDHybridNet Components. Comparative accuracy and F1-score of LDDHybridNet and six ablation variants to illustrate the contribution of various components to LDDHybridNet. Removing SE and CBAM modules results in a mild performance drop. At the same time, our model performs worse without ViT and ASPP, as those two components are critical in multi-scale and long-range feature modelling. However, removing temporal modules (such as Bi-LSTM and temporal attention) results in the most significant drop, confirming the importance of effectively capturing sequential patterns of lesion evolution to achieve high confidence in the final classification. Figure 23. Accuracy and F1-Score Comparison Across Ablation Variants. AUC scores of the complete LDDHybridNet model and its ablated variants were plotted in Fig. 24. The AUC is highest for the whole model, indicating the best discriminative ability. Lowering SE and CBAM cause modest decreases, and removing ViT and ASPP leads to lower AUC because their diary and multi-scale representations break down. The most significant drop occurs when we remove the Bi-LSTM or temporal attention modules, suggesting that temporal sequence modelling is required for consistent, high-confidence predictions under changing imaging conditions. AUC Comparison Across Ablation Variants. Robustness under field conditions To assess the robustness of LDDHybridNet, its ability to generalise to FieldPlant images collected in an uncontrolled agricultural setting with varying natural light conditions, complex background clutter, overlapping leaves, and natural occlusions was also tested. Even with these challenges, we achieved 95.37% accuracy, demonstrating that the model generalises well beyond the constrained PlantVillage environment. For most disease categories, performance degradation was minimal. However, there was some confusion among visually similar lesions, namely between early blight and late blight, under simulated low-light or high-shade conditions. For a more in-depth understanding of misclassified samples, we inspected them individually. We saw that errors generally occurred when the leaf was only partly visible due to background clutter (e.g., soil textures, crop residues) or surface properties (e.g., water droplets), or when a high gradient of illumination obscured the lesion’s structures. Finally, overlapping leaves or differences in angle introduce temporal inconsistency between pairs of ROIs and, in some cases, cause prediction uncertainty. These inconsistencies were avoided mainly by Bi-LSTM and attention-based temporal context modelling, which helped stabilise the feature representation across frames. We show the model’s predictions on a subset of the dataset and observe that disease-region localisation is challenging to perform correctly, as shown in the example fields and their predicted labels. One remarkable aspect we noticed from LDDHybridNet is that, despite the naturally complicated problem (i.e., LDD), the model can still extract the most relevant areas for a disease by utilizing ROI extraction and refining Grad-CAM attention, which suggests that the time-space integrated image recognition can provide substantial improvement in interpretability and validation of the decisions made. These results demonstrate that the framework maintains a high degree of performance under conditions closer to those encountered during caretaking tasks, as would be expected in typical precision agriculture scenarios. Early-stage disease detection analysis In precision agriculture, early-stage autonomous disease detection is essential for prompt intervention. LDDHybridNet is used for the detection of early-stage leaf diseases, which exhibit limited, localised, and mild visual symptoms (e.g., light discolouration, smooth-texture distortion, and low-intensity lesion patterns) and are naturally widespread before high-level infection. Due to its sensitivity to these faint symptomatic cues, our proposed model obtains an accuracy of 92.84% on a curated balanced subset of early-stage leaf images, significantly outperforming representative baseline models (over 5% for EfficientNet-B0 (86.31%), and over 7% for ViT-B/16 (84.92%), and CNN–LSTM hybrids (88.40%). Much of this improvement was due to the proposed ROI-guided pipeline. Even for small or imperfect disease lesions, isolating lesion-susceptible areas and suppressing background content made them easier for the CNN-LSTM encoder to interpret. The multi-scale ASPP module then boosts detection by capturing local texture differences, subtle venation distortions, and early necrotic gradients that traditional CNNs overlook. This was augmented by the ViT head to provide global contextual cues to distinguish an early blight–like anomaly from natural leaf pigmentation. Translations, Drag and Drop, and Upload Subtitle. Early Grad-CAM visualisation confirms that Grad-CAM provides additional interpretability than was observed earlier in the model. In contrast, while attention was often diffused across healthy tissue in baseline architectures, the attention map of LDDHybridNet consistently concentrated on early necrotic patches, small chlorotic halos, and low-contrast lesion boundaries. The motif of this focused activation pattern is consistent with expert annotations in agricultural literature, showing that the framework can effectively predict disease manifestation before harmful symptoms set in (often, these symptoms are only apparent to the human eye). This suggests that LDDHybridNet is effective for proactive precision agriculture workflows that aim to reduce crop loss by diagnosing plant disease early. Explainability and visual validation (Grad-CAM) To analyse the interpretability and diagnostic transparency of LDDHybridNet, Grad-CAM heatmaps were generated for (a) six disease categories (i.e. Tomato Mosaic Virus, Pepper Bell Bacterial Spot, Potato Early Blight, Potato Late Blight, and Tomato Target Spot) and (b) healthy leaf samples. The visualisation results show that, regardless of the specific symptom patterns, the model focused on biologically relevant regions of the lesions. Heatmaps for viral and bacterial infections, such as Tomato Mosaic Virus and Pepper Bell Bacterial Spot, also show patchy morbid lines, chlorotic specks, and tiny necrotic points, confirming the nuanced texture perturbations associated with these infections, so that the model does pick up morpho-spatial patterns related to these diseases. We found that Grad-CAM activation in Potato Early Blight and Late Blight corresponded to classic indicators of disease, such as circular rings, areas of water-soaking, and centres of dark brown lesions. The Heatmaps Tomato Target Spot highlight circular necrotic spots (orange) and halo-like yellow bands (green) near necrotic associations, which appear to align with agronomic symptom descriptions. In Fig. 25, healthy leaves can be clustered from diseased samples. Finally, healthy leaf Grad-CAM explanations display diffuse, low-intensity activations focused on venation rather than artifactual confounding indicative of model overfitting to non-disease cues. By contrast, in diseased leaves, the sharply bounded activation regions over pathological structures confirm that they actually extract net features that are aware of the lesions. Grad-CAM–Based Visual Explanations of LDDHybridNet Predictions Across Multiple Leaf Diseases and Healthy Samples. Both sides of the aisle — the right and the hard right — serve up examples aplenty. These misclassifications were primarily due to lesions with characteristics of more than one disease type or symptoms being obscured. To shed light on the reasoning behind these confusions induced by the model, we derived Grad-CAM maps, an example of which is presented in Fig. 4 (and more in the supplementary material), wherein the model almost exclusively draws attention to pixels with conflicting colour changes or overlapping textures through the two different phases of the disease resulting in the confusion between early blight and late blight. It allows error patterns to be interpreted, enabling practitioners to ensure that the model’s decision-making is based on biologically salient features rather than spurious artefacts. These visual evidences altogether assert that LDDHybridNet is highly explainable at a higher level, with attention patterns aligned with known plant pathology for the diseases studied. The overlap between predicted activation areas and actual lesion areas demonstrates that our model accurately predicts while reasoning in a clinically interpretable manner relevant to existing real-world precision agriculture workflows. 10 test for statistical significance/characterise replicability A statistical evaluation across multiple training scenarios was performed to test how consistent and reliable LDDHybridNet is. We used the 5-fold cross-validation protocol, and the results (Mean ± STD values of Accuracy, F1-Score, and AUC) are listed in Table 7 to improve reproducibility. This little variability between folds indicates excellent stability during training and low sensitivity to splitting. Table 7Statistical Reliability of LDDHybridNet Across 5-Fold Cross-Validation.MetricFold 1Fold 2Fold 3Fold 4Fold 5Mean ± SD95% CIAccuracy (%)98.7298.9199.0398.8099.0998.91 ± 0.14[98.73, 99.09]F1-Score (%)98.5298.6698.7998.6198.7898.67 ± 0.11[98.53, 98.81]AUC (%)98.9299.1099.2598.9799.3499.12 ± 0.16[98.92, 99.32] Statistical Reliability of LDDHybridNet Across 5-Fold Cross-Validation. Performance improvements over baseline models were assessed for statistical significance using paired t-tests and Wilcoxon signed-rank tests on fold-wise scores. They also outperformed all EfficientNet-B0 (p < 0.01), ViT-B/16 (p < 0.01), and CNN-LSTM baselines (p < 0.01), verifying that these gains were not random fluctuations across the tasks. Ultimately, these were obtained through very narrow 95% confidence intervals (CIs) and through this level of consistency - a mandatory prerequisite for real-world agricultural applications, where right, reproducible prediction is a must. In general, the statistical analysis shows that LDDHybridNet exhibits consistent, repeatable behaviour across different data splits, with statistically significant and practically relevant improvements. Boxplots of Accuracy, F1-Score, and AUC from 5-fold cross-validation to assess the stability and consistency of LDDHybridNet are depicted in Fig. 26. The first view shows that each metric has low variance (i.e., values across the different data splits are clustered closely together), indicating it is a reasonably consistent performer from split to split. Accuracy and AUC show very narrow ranges, indicating good generalisation and strong discriminative power. The consistent uniformity of the results across folds demonstrates that LDDHybridNet generates stable outputs independent of how the dataset was partitioned for training and testing, thereby providing further assurance of the model’s potential effectiveness for deployment in real agricultural contexts. Cross-Validation Performance Distribution of LDDHybridNet Across Accuracy, F1-Score, and AUC. Plot in Fig. 27 shows the statistical significances of the improvements of LDDHybridNet over three baseline models, pair-wise significance tests. 1. The −log10-transformed p-values are highlighted on the plot. Performance gains are very significant and not due to random variation over folds: all p-values are well below 0.01. Such statistical confidence for the presented analysis confirms that LDDHybridNet consistently outperforms EfficientNet-B0, ViT-B/16, and CNN-LSTM. Statistical Significance of LDDHybridNet Improvements Over Baseline Models Using Paired Tests. From the experimental results, it is observed that LDDHybridNet achieves high accuracy and robustness for leaf disease classification under both completely controlled and slightly rugged field conditions. Across all evaluation metrics, our ROI extraction, multi-scale feature encoding, temporal modelling, and attention mechanism consistently improve performance. Methods, such as class-wise analysis for visually similar diseases and specificity and sensitivity analyses for early-stage detection, reveal the lack of discriminability of our method and its excessive sensitivity to small lesion patterns. Statistical reliability tests provide evidence of consistent performance with low variance across folds, while Grad-CAM visualisations support attention that is consistent with biologically relevant features. In short, these outcomes validate the viability of LDDHybridNet as an interpretable and field-compatible precision agriculture solution. Discussion Despite the tremendous progress made by deep learning techniques for the most recent leaf disease detection works, the task of leaf disease detection in a real-world agricultural environment is still challenging due to a high level of intra-class variabilities, presence of subtle symptoms in initial stages, illumination variabilities, presence of cluttered backgrounds and domain shifts between laboratory and field images. Although CNN-based models tend to perform well on cleaned datasets, their generalisation ability drops significantly in unconstrained field——like conditions. Several Transformer-based and hybrid models partly mitigate scale variability but often do not incorporate explicit strategies to isolate biologically relevant regions in their interpretive outputs that implicate causative lesions, thus undermining their agronomic decision-support capacity. In this regard, LDDHybridNet demonstrates the ability to learn spatial constraints from ROI without visually distracting syntactic components, thereby increasing robustness, especially for weak-lesion early samples. The framework focuses feature extraction on leaf regions obtained via segmentation, which reduces background interference and is more sensitive to faint discolouration and localised texture distortions. Structured feature sequencing with Bi-LSTM-based contextual modelling enables the network to learn dependencies across spatial features, which is particularly useful when symptoms may be sparse or not visually separable. But the experimental results also show some significant failure modes, especially in harsh field conditions. This typically leads to performance degradation when early-stage lesions have a visual appearance indistinguishable from natural leaf texture variations, heavy occlusions obscure lesion boundaries, or when the chromatic cues are dominated by illumination artefacts. This may not isolate disease-relevant regions as fully as one would like, resulting in incomplete or noisy feature representations. The preprocessing step shows that there is still inherent sensitivity to image and segmentation quality in uncontrolled settings, despite the advantages of ROI-guided pre-processing. By progressively ablating combinations of components, the analysis indicates that the performance increase is caused by the components’ interactions in a mutually exclusive manner and is not due to any specific module. Though attention mechanisms and global context modelling help reduce discrimination, errors driven by visually confounding patterns (e.g., nutrient deficiencies or symptoms of abiotic stress) cannot be eliminated. As expected, the Grad-CAM visualisations show that, in most cases, the model attends to regions relevant to the lesion; however, being a post-hoc method, Grad-CAM provides only coarse localisation, which may not accurately localise features contributing to the prediction, especially in failure cases. In general, they seem to imply that LDDHybridNet has been solved in the field and is deployable, which is far from the truth and, in fact, the opposite; in fact, if we look at all the discussion of the results. Strengths are higher relative sensitivity over all early stages, interpretability, consistency across datasets; weaknesses highlight the need for more comprehensive datasets, more robust ROI, and complementary sensing modalities.” Limitations of the Study There are, however, some limitations to this study, despite the good overall performance. First, the evaluation is based on only two publicly available datasets (with limited geographic and crop diversity), which could potentially hinder generalisation to unseen fields. Second, the success of our approach relies heavily on accurate ROI extraction; strong occlusion, motion blur, or extreme lighting conditions may severely degrade segmentation quality and, in turn, classification results. Third, despite its improved interpretability, Grad-CAM offers post-hoc explanations rather than complete feature attribution or causal explanations. Lastly, no quantitative benchmarking for deployment on edge or resource-constrained platforms is provided, thereby limiting claims about real-time or field implementations. To overcome these limitations, it shall be necessary to use more extensive multi-region datasets, stronger segmentation strategies, explainability contrasts, and systematic evaluation on deployment-efficient hardware. These guidelines lay the foundation for future generalisations towards scaling up and building trustworthy AI in agriculture. Conclusion and future work In this paper, we introduce an ROI-aware CNN–LSTM hybrid framework, LDDHybridNet, for plant leaf disease detection, focusing on effective spatial feature representation learning, contextual dependency modelling, and explainability. Where ROI-guided preprocessing and structured feature modelling guided the approach, the engineered method efficiently directs focus to lesion-specific regions while limiting contamination from background noise, illumination variation, and partial occlusions, which are typical in agricultural images. Extensive exhaustive experiments conducted on the PlantVillage and FieldPlant datasets demonstrate that LDDHybridNet consistently outperforms all representative CNN-, transformer-, and hybrid-based baselines across various evaluation metrics, including accuracy, F1-score, and AUC, and with respect to its sensitivity to weak-lesion early-stage samples. Additionally, Grad-CAM–based interpretability analyses demonstrate that the model’s predictions are guided by biologically relevant lesion locations, thereby providing transparency and confidence in the diagnostic outputs. The results show that the framework proposed here provides a reliable, interpretable method for automatically analysing leaf disease in precision agriculture pipelines. Although these results are promising, the present study is not without limitations. The framework has been evaluated on a small number of datasets with limited geographic and crop workflow diversity and is only as effective at ROI extraction in challenging imaging environments. In addition, this work did not investigate the generation of deployment benchmarks for edge or embedded platforms, although the architecture is intended for computationally constrained environments. Future work will extend this approach to broader validation using multi-regional field datasets covering different climate and crop species contexts. The generalisation to heterogeneous agricultural environments may also be improved by extending the framework through domain adaptation, self-supervised learning, and meta-learning. Another potential avenue to enhance early disease detection is to include additional sensing modalities, like hyperspectral imagery, Unmanned Aerial Vehicle-based data or environmental metadata. Last, it will be essential to investigate alternative explainability methods beyond post-hoc Grad-CAM and to evaluate the effectiveness of lightweight, deployment-oriented implementations so that agricultural intelligence systems can be scalable and practically implementable. Author contributions All authors contributed to the study’s conception and design. Material preparation, data collection, and analysis were performed by Ch. Santoshi, Suresh kumar mandala, L Raghavendar Raju.The first draft of the manuscript was written by CH. Santoshi all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript Funding No financial support was received by the authors in this research. Data availability Data is available with the corresponding author and will be given on request. Code availability The code is available from the corresponding author and can be given on request. Materials availability Materials used in this research are available with corresponding author and given on request. Code availability and reproducibility Due to institutional and collaborative constraints, the full source code for LDDHybridNet cannot be released publicly at this time. To facilitate reproducibility, the manuscript provides a detailed algorithmic description, preprocessing pipeline, configuration parameters, and training protocol. Pretrained model weights and supporting scripts can be made available to reviewers or researchers upon reasonable request. Public release of the code will be considered in future once existing restrictions are resolved. Declarations Competing interests The authors declare no competing interests. Ethical approval This research does not involve humans or animals, so no ethical approval is required. Conflict of interest LDDHybridNet: An ROI-Aware CNN–LSTM Hybrid Framework for Accurate and Early Leaf Disease Detection in Precision Agriculture. The authors whose names are listed immediately below certify that they have NO affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers’ bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or non-financial interest (such as personal or professional relationships, affiliations, knowledge or beliefs) in the subject matter or materials discussed in this manuscript. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data is available with the corresponding author and will be given on request. The code is available from the corresponding author and can be given on request. Materials used in this research are available with corresponding author and given on request. Due to institutional and collaborative constraints, the full source code for LDDHybridNet cannot be released publicly at this time. To facilitate reproducibility, the manuscript provides a detailed algorithmic description, preprocessing pipeline, configuration parameters, and training protocol. Pretrained model weights and supporting scripts can be made available to reviewers or researchers upon reasonable request. Public release of the code will be considered in future once existing restrictions are resolved.