Design of an AI-Driven Analytical Framework Integrating Machine Learning and Hyperspectral Remote Sensing for Detection and Classification of Locust-Prone Areas in South Nyanza Kenya

Desert locust outbreaks pose a persistent threat to agricultural productivity and food security in East Africa, while conventional surveillance approaches remain limited by delayed reporting, restricted spatial coverage, and weak predictive capability. This study develops and evaluates an artificial intelligence-driven analytical framework that integrates machine learning with hyperspectral remote sensing for detecting and classifying locust-prone areas in South Nyanza, Kenya. An empirical quantitative experimental design was applied to 150 georeferenced spatial observation units using spectral, vegetation, and bioclimatic indicators derived from Sentinel-2, CHIRPS, and MODIS data. The analytical framework incorporated automated feature engineering, Random Forest classification, Logistic Regression, and geospatial hazard visualization. The Random Forest model, configured with 500 decision trees and an mtry value of 3, achieved an overall classification accuracy of 81.33%, precision of 81.58%, recall of 81.58%, F1-score of 81.58%, and an Out-of-Bag error rate of 18.67%. The validated ROC-AUC reached 0.835, indicating good discrimination capability. Variable importance analysis identified precipitation, soil moisture represented by SAVI, and vegetation greenness represented by NDVI as the most influential predictors. Logistic Regression showed a positive association between hyperspectral indicators and locust-prone classification, although the predictors were not statistically significant at the 5% level. The findings demonstrate that integrating machine learning with remotely sensed environmental indicators provides a scalable approach for strengthening locust surveillance and supporting evidence-based early warning systems.