Multispectral Image Fusion for Agriculture

Abstract

Rice-leaf diseases such as Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf scald, and Sheath Blight reduce crop yield and are conventionally diagnosed by slow, expertise-dependent manual inspection. This paper presents a feature-based rice-leaf disease classification system that compares an RGB-only pipeline (18 features: RGB, HSV, and GLCM texture statistics) against an RGB + Pseudo-NIR feature-level fusion pipeline (32 features, additionally incorporating Pseudo-NIR statistics, Pseudo-NIR-derived NDVI descriptors, and Pseudo-NIR GLCM texture). Both vectors are standardized and classified with a Gradient Boosting classifier under an identical configuration for both pathways. Trained and evaluated on 3,829 labelled rice-leaf image pairs spanning six classes (2,680 training, 574 validation, 575 test), the RGB-only model reached 83.13% test accuracy and the fusion model reached 85.74%, a 2.61 percentage-point improvement corroborated by 5-fold cross-validation (82.64% ± 1.39%) but not reaching conventional statistical significance under a two-proportion z-test (z = 1.22, p = 0.22). The classifier is deployed as a full web application with user authentication and a rice-leaf validation guard that rejects non-rice-leaf images before they reach the classifier. The contribution is therefore twofold: a controlled, identically-configured RGB-only vs. fusion ablation, and a fully deployed, explainable, GPU-free system built around it.