<p>Outbreaks of the larch caterpillar (<i>Dendrolimus superans</i>) cause severe ecological and economic damage to boreal forests, underscoring the urgent need for effective monitoring and early warning systems. However, the utility of space-borne multispectral imagery (MSI) for this purpose is often constrained by either coarse spatial resolution or insufficient spectral bands, limiting the accurate classification of pest occurrence levels. To address this challenge, we developed an NDVI-constrained Dynamic Ridge Polynomial Neural Network (NDRPNN) to fuse Sentinel-2 MSI data with Gaofen-2 (GF-2) panchromatic imagery, thereby enhancing spatial detail while preserving spectral integrity. Time-series spectral, textural, and polarimetric features derived from Sentinel-1/2 imagery were subsequently integrated, and correlation analysis was applied to identify the most sensitive indicators. Four classification models—Random Forest, Light Gradient Boosting Machine, Stacking Ensemble, and Soft Voting Ensemble (SVE)—were evaluated for detecting infestation levels, with Shapley (SHAP) analysis employed to interpret feature contributions. The NDRPNN exhibited robust fusion performance in forested landscapes. Ensemble methods outperformed single classifiers, with the SVE model achieving the highest accuracy (overall accuracy = 87.6%, Kappa = 0.83). SHAP analysis identified the mean and maximum Normalized Difference Vegetation Index (NDVI), minimum Anthocyanin Reflectance Index (ARI), minimum Normalized Burn Ratio (NBR), and seasonal amplitude of Enhanced Vegetation Index (EVI) as key contributing features, highlighting the critical role of time-series vegetation indices and textural metrics in early pest detection. This study demonstrates that the integration of high-quality Sentinel-2 and GF-2 imagery with ensemble learning enables rapid and precise assessment of pest occurrence, offering a robust foundation for the early warning and scientific management of forest pests in mountain regions.</p>