Lightweight Multimodal Rice Leaf Disease Screening and Nepal-Localized Climate-Smart Advisory Support: Evaluations on Real Rice Images, Nepal SOTER Soil Profiles, and WorldClim Climate Data

Abstract

This study evaluates a lightweight multimodal pipeline for rice leaf disease screening and Nepal-localized climate-smart advisory support. The revised evaluation uses directly downloadable public datasets rather than procedurally generated image or soil-weather records. The vision branch uses the UCI Rice Leaf Diseases dataset, which contains 120 real JPG images in three balanced classes: bacterial leaf blight, brown spot and leaf smut. The environmental branch combines Nepal SOTER soil profile data with WorldClim 2.1 monthly climate rasters for precipitation, minimum temperature, maximum temperature, water vapor pressure and elevation, producing 645 profile-season records from 129 georeferenced Nepal soil profiles. The advisory branch uses 24 curated English-Nepali rice management cards and 96 farmer-style retrieval queries grounded in Nepal rice disease, soil fertility and crop-management references. Five image classifiers were compared using 136 engineered color, lesion and edge features. Logistic regression achieved the strongest clean hold-out performance, with 0.917 accuracy and 0.918 macro-F1. Stress tests showed that the model is reasonably stable to small rotations and moderate blur, but accuracy declines under strong sensor noise and large brightness shifts. In the advisory retrieval experiment, character TF-IDF performed best with 0.875 top-1 accuracy, 0.969 top-3 accuracy and 0.922 mean reciprocal rank. For the soil-climate risk task, random forest achieved the best hold-out macro-F1 of 0.869 and cross-validated macro-F1 of 0.851. The results support the paper's central claim: a small, inspectable system can connect leaf-image evidence, Nepal soil-climate context and grounded farmer advice, while remaining explicit about the limits of non-field diagnostic evaluation.