<p>Agroecological zoning (AEZ) frameworks employ two primary methodological approaches: knowledge-driven Multi-Criteria Decision Analysis (MCDA) and data-driven machine learning algorithms, each offering distinct advantages for spatial delineation. Despite their complementary nature, these methods have been applied separately, lacking systematic integration to leverage their combined strengths in data-scarce tropical regions. To address this, we developed a novel hybrid machine learning framework integrating Boruta feature selection, expert-refined Analytical Hierarchy Process (AHP) weighting, and K-means clustering, applied to South-Kivu Province, eastern Democratic Republic of Congo (DRC). The framework analyzes 18 biophysical predictors spanning climatic, topographic, vegetation health, land cover, and soil properties derived from satellite observations at 30-m spatial resolution. Derived AEZs were validated against rule-based simulated traditional zones using Random Forest (RF) classifier on 42,880 stratified points with 70:30 train-test partitioning. Boruta objectively confirmed all factors as relevant with the length of the growing period (LGP) as the most important factor (relative importance = 52.04). Clustering delineated six AEZs with contrasting agricultural potential and management priorities. Validation achieved 91.5% overall accuracy, Kappa of 0.896, and per-class AUC values &gt;0.986. This hybrid framework advances agroecological mapping by integrating feature optimization, expert knowledge, and quantitative validation, overcoming interpretability-precision trade-offs of standalone approaches. The resulting high-resolution AEZs provide a scalable blueprint for climate-smart agricultural planning and targeted land management in heterogeneous, data-limited tropical ecosystems.</p> Graphical Abstract <p></p> <p>Based on graphical snapshot, this study presents a novel hybrid machine learning framework for high-resolution agro-ecological zoning in the South-Kivu province, eastern Democratic Republic of Congo. Eighteen biophysical factors derived from multiple satellite and global datasets (SRTM for topographic, HWSD and SoilGrids for soil properties, ESA-WorldCover for land cover, MODIS for vegetation health, IMERG for precipitation, and CPC for temperature) were integrated at 30-m resolution. These span climatic (Length of the growing period (LGP), precipitation, temperature, rainfall seasonality index (RSI), temperature seasonality index (TSI), ardity index (AI), topographic (elevation, slope, landform, aspect), vegetation (vegetation health index (VHI), land cover), and soil (cation exchange capacity (CEC), bulk density, pH, soil type, soil texture, organic matter) variables. The workflow comprises four sequential steps. First, the Boruta feature selection objectively identifies and retains the most relevant predictors, minimizing noise and overfitting. Second, Random Forest (RF) derived initial importance is refined by 12 field expert consultation through two rounds Delphi-inspired process to inform the Analytical Hierarchy Process (AHP), yielding differential weights with consistency checks (CR &lt; 0.10). Weighted factors are normalized (min-max for positive and inverse for negative relationships) and subjected to K-means clustering (optimized via the elbow method, nstart = 25), with a water mask applied using elevation and slope thresholds to exclude non-agricultural areas. Resulting AEZs are validated against simulated traditional zones using Random Forest classification on 42,880 stratified samples (70% training, 30% testing), with 10-fold cross-validation on training data. The framework delineates six spatially explicit AEZs across South Kivu’s heterogeneous landscape, shown in the color-coded map distinguishing zones from highlands to lowlands with contrasting agricultural potential. LGP, precipitation, temperature, RSI, and aridity index were the most contributors (collectively explaining 45% of zonal variation). Validation confirms outstanding performance (accuracy 91.5%, Kappa 0.896, per-class AUC &gt; 0.986), providing reproducible, high-resolution boundaries for targeted, climate-smart agricultural planning and sustainable land management in tropical data-scarce environments.</p>