AI-DRIVEN SMART FARMING: FROM CROP PREDICTION TO PLANT DISEASE DETECTION

Abstract

The escalating challenges to global food security, driven by climate change, limited natural resources, and rising population demands, necessitate the adop tion of intelligent, data-driven solutions in agriculture. This thesis presents two complementary deep learning frameworks aimed at addressing key aspects of precision farming: accurate pre-season crop prediction and robust plant disease detection. For crop prediction, a hybrid model integrating Convolutional Neu ral Networks (CNNs) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks is developed, effectively capturing spatial patterns and temporal trends from historical crop rotation data and synthetic field-level features. To ensure generalizability and prevent overfitting, the model leverages stratified k-fold cross validation and dropout regularization, consistently outperforming conventional methods in terms of predictive accuracy and applicability to real-world scenarios. In parallel, this work introduces a novel Vision Transformer (ViT) combined with a modified High-Resolution Network (HRNet) for disease diagnosis across multi ple plant species, addressing challenges such as variation in leaf venation, texture, and symptom presentation. By fusing global contextual reasoning from ViT with f ine-grained spatial precision from HRNet, the proposed architecture achieves su perior classification accuracy in both controlled and field environments. Together, these models provide an end-to-end framework for predictive and preventive crop management, advancing the goals of sustainable agriculture, early intervention, and resilient food systems.