Search

The potato (Solanum tuberosum L.) is one of the most important food crops and is consumed by more than a billion people worldwide. Despite the fact that crop growth models are widely used to optimize crop management and agricultural decision-making, they fail to provide spatial information, whereas satellite-based methods do, and in a cost-effective manner. In order to improve satellite-based crop yield prediction models, we developed a methodology based on Sentinel-2 (S2) data using Machine Learning (ML) techniques. In addition, we included and compared the commonly used Normalized Difference Vegetation Index (NDVI) to a newly developed spectral indicator, the Potato Productivity Index (PPI). Our results showed the capacity of our ML models to predict potato yield with great accuracy (coefficient of determination (R2) = 0.77 and Root Mean Square Error (RMSE) = 15.42%), using S2 bands 2, 3, 4, 5, 8A, and 9, as well as the PPI index. This index aims to provide information concerning plant photosynthetic activity as well as its water stress. The Random Forest (RF) model using the S2 bands and the PPI index as predictors obtained better predictive results (R2 = 0.77, RMSE = 15.42%) for potato crops than using S2 bands and NDVI (R2 = 0.66, RMSE = 16.88%). This study demonstrates the suitability of our models to predict potato yields in the region studied and improves on a previous approach presented by the same authors in terms of model interpretability and inter-annual variability. [ABSTRACT FROM AUTHOR]

Crop growth modeling and yield forecasting are essential to improve food security policies worldwide. To estimate potato (Solanum tubersum L.) yield over Mexico at a municipal level, we used meteorological data provided by the ERA5 (ECMWF Re-Analysis) dataset developed by the Copernicus Climate Change Service, satellite imagery from the TERRA platform, and field information. Five different machine learning algorithms were used to build the models: random forest (rf), support vector machine linear (svmL), support vector machine polynomial (svmP), support vector machine radial (svmR), and general linear model (glm). The optimized models were tested using independent data (2017 and 2018) not used in the training and optimization phase (2004–2016). In terms of percent root mean squared error (%RMSE), the best results were obtained by the rf algorithm in the winter cycle using variables from the first three months of the cycle (R2 = 0.757 and %RMSE = 18.9). For the summer cycle, the best performing model was the svmP which used the first five months of the cycle as variables (R2 = 0.858 and %RMSE = 14.9). Our results indicated that adding predictor variables of the last two months before the harvest did not significantly improved model performances. These results demonstrate that our models can predict potato yield by analyzing the yield of the previous year, the general conditions of NDVI, meteorology, and information related to the irrigation system at a municipal level. [ABSTRACT FROM AUTHOR]