SCALING SUBSURFACE YIELD PREDICTION: MACHINE LEARNING AND UAV-SATELLITE DATA FUSION FOR ROOT CROP AGRONOMY

Abstract

Root and tuber crops including potato (Solanum tuberosum L.), cassava (Manihot esculenta Crantz), sugar beet (Beta vulgaris L.), peanut (Arachis hypogaea L.), and sweet potato (Ipomoea batatas L.) underpin global food and feed security, yet their harvestable organs develop below ground and remain invisible to optical remote sensing throughout the growing season. This review brings together evidence from approximately ninety peer-reviewed studies to evaluate the state of UAV-satellite data fusion combined with machine learning (ML) for non-destructive subsurface yield prediction. We trace the agronomic logic by which above-ground canopy traits,leaf area index, chlorophyll, canopy temperature, and structural height-serve as physiological proxies for tuber and root bulking, and document how spatiotemporal trade-offs between UAV (centimetre-scale, campaign-based) and satellite (decametre-scale, regular revisit) sensing are reconciled through pixel-, feature-, and decision-level fusion. Empirically, model accuracy spans R² ≈ 0.5–0.87 across crops, platforms, and ML architectures, with the strongest results reported for cassava on UAV derived canopy height (R² = 0.87) and for potato on Sentinel-2 (Random Forest R² = 0.77 outperforming SVM R² = 0.66), while peanut, sugar beet, sweet potato, and neglected root crops such as taro and tiger nut typically achieve R² between 0.5 and 0.85. Multi-task deep learning architectures and cross-crop transfer learning now promise generalisable subsurface yield models, and integration of ground-penetrating radar, environmental covariates, and geostatistical interpolation with canopy remote sensing is emerging as the route to scalable regional decision support. We conclude by identifying the operational barriers-computational cost, cultivar specificity, and the absence of foundation models for root crop agroecosystems-that must be overcome to translate these proof-of-concept accuracies into operational agronomic use.