Optimizing potato yield mapping and prediction: Integrating satellite-based remote sensing and machine learning for sustainable agriculture

Precision agriculture and sustainable farming require crop yield prediction and mapping. PEI is a major Canadian potato producer. However, PEI potato yield prediction research is limited. This highlights a literature gap and the need for improved data-driven precision agriculture in PEI. High-resolution satellite imagery and machine learning (ML) enable field-scale crop yield mapping. This study investigated the potential of high-resolution multispectral imagery and ML for potato yield prediction. The study focused on four plots in PEI during the 2021 and 2022 growing seasons. Potato crop yield data collected using a combined harvester and manual digging were analyzed to model yield using Sentinel-2A and PlanetScope imagery. For both sensors, vegetation indices (NDVI, GNDVI, SAVI, and EVI) and spectral bands chosen for their application in crop growth monitoring were retrieved and incorporated into ML models. The cloud computing platform, Google Earth Engine (GEE), was used to evaluate and compare the performance of three ML algorithms, which are random forest regression (RFR), classification and regression trees (CART) and gradient tree boosting (GTB). Overall, the performance of all three models was satisfactory in yield prediction with both sensors. However, GTB with Sentinel-2A and harvester data gave slightly higher estimation accuracy with R² values of 0.71–0.78, RMSE values of 2.82–5.96 t/ha, and MAE values of 2.33–4.2 t/ha. Hence, the approach used in this study provides real-time seasonal yield prediction maps on a field scale, which will help the farmers identify the targeted areas for variable rate application, leading to resource efficiency and sustainability.