Deep Hybrid Learning for Smart Agriculture using CNN-based Feature Extraction and LSTM-BiLSTM Sequence Modeling for Robust Plant Disease Detection

Plant diseases significantly reduce global crop productivity, creating an urgent demand for intelligent, automated diagnostic systems in agriculture. Traditional manual inspection is labor-intensive, subjective and often ineffective in detecting early or latent symptoms. This study presents a multi-class classification and severity estimation framework for ten plant disease categories: Maize brown spot, maize rust, maize healthy, potato early blight (Alternaria solani), potato late blight (Phytophthora infestans), potato healthy, soybean mosaic virus (SMV), soybean pod mottle virus (SPMV), soybean sudden death syndrome (SDS/SBS) and soybean healthy. The objective is to develop a robust hybrid deep learning model capable of accurate early detection and quantitative severity assessment to support precision agriculture.

A hybrid architecture combining convolutional neural networks (CNN) with LSTM and BiLSTM networks was implemented. The preprocessing pipeline included leaf segmentation, binary masking, defect localization and edge detection to enhance lesion visibility. CNN layers extracted spatial and textural features, while recurrent layers modeled contextual dependencies within feature representations. Performance was evaluated using Precision, Recall, F1-score, defect percentage estimation, convergence analysis and t-SNE visualization.

Results demonstrated stable convergence with decreasing loss (0.8-1.2) and improved feature clustering. Defect severity ranged from 0.00% (Soybean healthy) to 87.93% (Maize brown spot). The framework enables early detection (0.29-5% infection), reduces yield loss, minimizes chemical overuse and promotes sustainable smart agriculture systems.