Wheat Yield Prediction with Machine Learning based on MODIS and Landsat NDVI Data at Field Scale

Accurate estimation of wheat yield using Remote Sensing-based models is critical in determining the effects of agricultural drought and sustainable food planning. In this study, Winter wheat yield was estimated for large fields and producer fields by applying Normalized Difference Vegetation Index (NDVI) based linear models (simple linear regression and multiple linear regression) and Machine Learning (ML) techniques (support vector machineₛvm, multilayer perceptronₘlp, random forestᵣf). In this study, depending on the ecological zone, crop sampling was carried out from 380 rainfed parcels where wheat was planted. On the basis of crop development periods (CDP), the highest correlation between NDVI and yield occurred during the flowering period. In this period, coefficient of determination (R2) was 63% in TIGEM fields and 50% in producer fields for MODIS data, and 61% and 65% for Landsat data, respectively. In TIGEM fields, the best prediction performance was obtained with the MLP model for MODIS (RMSE: 0. 23-0. 65 t/ha) and Landsat (RMSE: 0. 28-0. 64 t/ha). On the other hand, the highest forecasting accuracy was acquired with the SVM model in producer fields. The RMSE values ranged from 0. 74 to 0. 80 t/ha for MODIS and 0. 51 to 0. 60 t/ha for Landsat 8. The error value obtained with MODIS was approximately 1. 4 times higher than the Landsat 8 data in producer fields. For yield estimation, the best estimation can be made 4-6 weeks before the harvest. In regional yield estimations, satellite-based ML techniques outperformed linear models. ML models have shown that it can play an important role in crop yield prediction. In crop yield estimation, it is a priority to consider the impact of climate change and ecological differences on crop development.