High‐resolution satellite imagery and machine learning for predicting yield and nitrogen uptake in potatoes

Optimizing nitrogen (N) management is crucial for maximizing potato ( Solanum tuberosum L.) yield while minimizing environmental impacts. Achieving this requires timely, in‐season estimation of crop N status and the ability to predict final tuber yield early in the season. However, commonly used approaches, including destructive sampling or UAV‐based sensing, are often labor‐intensive, costly, and limited in scalability. This study evaluated the utility of high‐resolution SkySat satellite imagery combined with machine learning (ML) to predict final tuber yield and in‐season N uptake within five nutrient management experiments in northeast Florida. Vegetation indices (VIs) derived from satellite imagery were used to train three ML algorithms: random forest (RF), support vector regression (SVR), and partial least squares regression. Under nested cross‐validation, models achieved high predictive accuracy for yield ( R 2 = 0.74–0.79), with RF performing well up to 77 days after planting ( R 2 = 0.78, normalized root mean square error [nRMSE]% = 16.1%), and SVR enabling early‐season predictions before the final N application window ( R 2 = 0.76, nRMSE% = 16.5%). SkySat‐derived VIs also effectively estimated mid‐season N uptake ( R 2 up to 0.82). SHapley Additive exPlanations analysis identified modified soil‐adjusted vegetation index, soil‐adjusted vegetation index, and Thiam's transformed vegetation index as key predictors. However, when evaluated under a leave‐one‐experiment‐out framework, model performance declined substantially, with negative R 2 values, indicating limited generalization across experiments. These findings demonstrate that while high‐resolution satellite imagery combined with ML can provide accurate in‐season predictions within a given experimental setting, model transferability across environments remains a critical limitation. The results highlight the need for multi‐environment datasets and robust modeling approaches for precision agriculture applications.