Edge-Enabled Deep Learning and Multi-Model Hybrid Intelligent Architectures Integrating Sentinel-3 and MODIS Data for Real-Time Monitoring of Crop Stress and Nonlinear Pest Complexes Forecasting Under Climate Shifts

A twelve-year (2011–2022) multi-site agro-ecological study on the population dynamics of 5 key lepidopteran and hemipteran pests of pigeonpea ( Cajanus cajan ) in a sub-continental tropical environment was carried out to understand vulnerability patterns of this important pulse ecosystem to multiple biotic stresses. Twelve hundred and fifty hectares under diverse agro-ecological zones were surveyed using sub-meter accuracy real-time-kinematic global navigation satellite system coordinates and UTM projections. Using MODIS, Sentinel-3, and WorldClim datasets, drift-adjusted geospatial rasters were generated to represent the pest’s microclimatic and topographic preferences while overcoming local-scale heterogeneity. Analysis over standard meteorological weeks SMW 1–52 revealed that regional pest outbreaks followed closely the host plant phenology, while peaks in economic damage were strongly correlated to post-monsoon relative humidity gradients. Spatial autocorrelation estimates using Global Moran’s I: 0.65–0.82, p < 0.001, and surface interpolation using Universal Kriging corroborated a range of spatial dependence between 9.5 and 15.2 km and an associated cross-validation root mean square error of 0.05–0.14. With respect to benchmarking against state-of-the-art machine learning and deep learning algorithms, the proposed hybrid Fuzzy Probabilistic Neural Network demonstrated outstanding approximation capability with probabilistic layers based on Gaussian mixture density functions h = 0.5, m = 2.0, and achieving coefficients of determination between 0.961–0.991 as compared to Vision Transformers: 0.948–0.979, XGBoost: 0.889–0.935 and native SARIMAX: 0.810–0.872. With respect to the predictive performance, Monte Carlo dropout simulations between ∓ 0.028 and ∓ 0.045 and conformal prediction intervals with over 95% empirical coverage established the algorithm’s predictive reliability under stress conditions. Finally, with the deployment of the hybrid network’s ONNX runtime quantized variant achieving a 4.4× acceleration in inference, the methodology provides critical decision support to promote sustainable smallholder agrarian livelihoods and food security while enabling adaptive capacity to local-level biophysical and climatic heterogeneity.