Performance Evaluation and Validation of an AI-Driven Hyperspectral Remote Sensing Framework for Locust Surveillance Using Ground-Truth Data

Desert locust outbreaks continue to threaten agricultural productivity and food security across East Africa, creating an urgent need for surveillance systems that are accurate, scalable, and capable of supporting early intervention. This study evaluates and validates an AI-driven hyperspectral remote sensing framework for locust surveillance in South Nyanza, Kenya, using ground-truth observations and conventional field-scouting records as benchmarks. The study employed 150 georeferenced spatial units, with 70% used for model training and 30% reserved for independent validation. Random Forest and Logistic Regression were applied to hyperspectral and environmental indicators, while model performance was assessed using confusion matrix metrics, Out-of-Bag error, and Receiver Operating Characteristic analysis. On the 45-unit validation set, the AI-driven framework correctly classified 37 locations, achieving an accuracy of 82.22%, precision of 83.33%, recall of 83.33%, and an F1-score of 83.33%, with an Out-of-Bag error rate of 18.20%. The Random Forest model achieved an ROC-AUC of 0.844, substantially higher than the 0.585 obtained from the conventional field-scouting baseline. The framework also reduced false-negative detections from 10 to 4 cases. These findings demonstrate that integrating machine learning with hyperspectral remote sensing can strengthen locust surveillance by improving classification reliability, reducing missed infestations, and supporting evidence-based early warning and intervention planning.