HYBRID DEEP LEARNING AND EXPLAINABLE AI FRAMEWORKS FOR PREDICTIVE ANALYSIS AND DECISION SUPPORT IN AGRICULTURE AND MENTAL HEALTHCARE

Abstract

Agricultural sector remains a cornerstone for boosting the economy of country, uplifting socio economic development and leveraging livelihoods of underprivileged society. However, crop diseases remain a major threat to this sector. Past studies have highlighted about ML and DL model deployment and have even emphasized on complex architecture to bridge lab-to field gap, but, on specific crops or in assessing the performance drop. This study uses a hybrid architecture composed of an EfficientNet (variant B0) backbone as a feature extractor and a linear classification head across cross crop datasets, comprising lab-style images of Tomato, Potato and Bell Pepper, and real field leaves images of Chilli and Brinjal, encompassing 22,184 curated images. The essence of this research lies in implementing a hybrid, light-weight, computationally efficient architecture for cross-crop diseases detection through leaf images obtained from heterogeneous environments. The deployed model achieved an average accuracy of 99.8 % on controlled datasets and 97.35% on real field datasets. The novelty of this research lies in its implementation and practicality to achieve notable performance, by comparing performance across cross crop dataset of diverse crops. By utilizing EfficientNet-B0 backed by linear classification head, this study avoids architectural complications and achieves notable performance, emphasizing deployment on-device applications in precision agriculture.