Adaptive Echo State Network for crop yield prediction incorporating Fall Armyworm dynamics

• Developed a novel Echo State Network model integrating Fall Armyworm (FAW) infestation data with a biologically informed penalty function for maize yield prediction • Achieved 67% reduction in prediction errors compared to traditional models, with R² of 0.55 during FAW outbreaks • Demonstrated accurate yield loss forecasting during severe pest outbreaks, closely aligning with observed losses exceeding 20% • Combined IoT sensor networks with machine learning to create an early-warning system for precision pest management • Established a semi-interpretable AI approach that bridges data science and agronomic knowledge for sustainable agriculture • Enables transition from reactive, chemical-intensive pest control to proactive, targeted interventions reducing environmental impact Agricultural productivity worldwide is threatened by invasive pests, notably the Fall Armyworm (FAW, Spodoptera frugiperda ), which has devastated maize yields across Africa and Asia since 2016. To support precision pest management, we developed an adaptive Echo State Network (ESN) that predicts annual maize yield while accounting for FAW pressure. We compiled a 15-year (2010–2024) monthly dataset combining satellite vegetation indices, in-field weather, soil chemistry readings, and FAW surveillance counts. FAW severity is quantified on a 0-100 scale by blending trap counts (40 %) and larval density (60 %) per month. First, the ESN is trained on all available data to predict crop yields based on environmental features. We then apply isotonic regression to map pest infestation levels to the ESN's residual over-predictions, producing a monotonic penalty curve. This curve quantifies yield losses at different pest pressures. During prediction, we apply this learned penalty to the raw ESN output, adjusting yield estimates to account for pest damage without altering the original ESN model. In cross-validation, the FAW-aware ESN achieves an R² of ∼0.55 and reduces prediction errors by up to 67 % versus unpenalized baselines, closely capturing observed yield reductions exceeding 20 % during severe outbreaks. The model outperforms standard regression and deep neural network approaches by similar margins. It guides farmers in targeting interventions to high-risk zones, reducing pesticide use and operational costs. These results highlight its value as an early-warning tool for targeted interventions that minimize chemical inputs and optimize resource allocation. Ongoing field validations will evaluate its scalability and practical impact in FAW-affected maize production regions.