A DUAL-HEAD CNN-LSTM FRAMEWORK FOR SIMULATING AND ANALYZING PLANT DISEASE PROGRESSION USING SYNTHETIC TEMPORAL SEQUENCES

In this study, a novel deep learning framework for plant disease analysis that simulates disease progression using synthetic temporal sequences generated from static leaf images is proposed. Traditional image-based disease classifiers ignore the temporal nature of plant pathology, which limits interpretability and prediction accuracy. To address this, the sequences are synthesised by interpolating between a pseudo-healthy image and a diseased sample using alpha blending. These sequences mimic gradual progression across four frames. The design of a dual-head CNN-LSTM model that takes these sequences as input, leveraging convolutional layers to extract spatial features and LSTMs to capture temporal dynamics. The architecture branches into two heads: one for disease classification via softmax and another for severity estimation through regression. Further integrated a frame-wise attention mechanism and temporal consistency regularisation to improve interpretability and stability. Evaluation on the corn leaf disease dataset demonstrates strong performance, achieving 92.16% classification accuracy and a severity MSE of 0.0001. The comparison of the proposed model against traditional classifiers like Random Forests and Logistic Regression shows a 7–10% improvement in F1-score and significantly better severity awareness. Our approach offers a temporally aware, explainable AI solution for agricultural monitoring. Severity curves and attention maps enable not only post-hoc analysis but also facilitate real-time field-level diagnostics. The pipeline's ability to generate balanced data across underrepresented disease classes is crucial in imbalanced datasets. Moreover, the use of synthetic progression opens opportunities for applying this model to other domains lacking temporal labels. The modular structure of the system supports easy plug-and-play integration with UAVs, mobile apps, and remote sensing platforms. In conclusion, this work bridges the gap between image classification and temporal disease modelling, making a strong case for progression-aware AI in digital agriculture.