Explainable Fertiliser Recommendation and Disease-Agnostic Severity Estimation Framework

Precision agriculture entails smart decision-support systems that are precise and can be used for nutrient management and plant health monitoring. The existing approaches are inclined to consider fertiliser recommendation and plant disease evaluation to be two distinct processes, but the analysis of plant diseases is primarily focused on classification and does not give much care to the extent of infection. Besides this, most severity estimation methods for segmentation or pixel annotations are based on learning and, thus, are hard to apply in practice. This paper presents an interpretable and modular system of agricultural decision support, which contains two independent systems: an explainable fertiliser recommendation system and plant disease detection with a training-free severity estimation pipeline. The MobileNetV2 transfer learning model is applied in order to first of all classify the plant diseases using a large multi-class dataset of plant diseases. A disease-agnostic Disease Severity Index (DSI) is then used to determine the severity of the disease using classical image processing features, which include lesion area ratio, colour deterioration, and shape irregularity. Fertiliser recommendations are done with the random forest classifier that has in-built tree-based explainability. The conducted findings of the experiments indicate that reliable disease detection, assessment of constant severity, and precise and interpretable fertiliser advice exist and support the feasible and data-efficient use of precision agriculture.