Remote sensing data and machine learning to predict yields of major crops on regional scale

Abstract

Pavlo Lykhovyd, Raisa Vozhehova, Liudmyla Hranovska, Oleksandr Averchev, Anatolii Tomnytskyi, Nataliia Avercheva, Mariia Nikitenko, Serhii Haievskyi, Oleg Shalar

Remote sensing and machine learning tandem is a powerful tool for crop monitoring and yield prediction. Current study is devoted to the evaluation of the relationship between five remote sensing indicators, affecting crop yields, such as NDVI, NDMI, VHI, LST and PET, as well as establishing the connectivity of these indices with the yields of twelve major crops cultivated in Ukraine during the period 2015-2023. Yielding data were retrieved form official statistical bodies of Ukraine. Remote sensing data were calculated and generalized through the Google Earth Engine platform through the requests to the API in JavaScript. Correlation and regression analysis were performed using common methodologies, as well as more robust machine learning techniques like Random Forest and Gradient Boosting Regression were also applied for yield prediction. It was determined that the strongest correlation (0.93-0.96) is between such indices as NDVI and VHI, LST and PET, VHI and PET. As for crop models, for most crops Random Forest algorithms provided the best overall quality of fitting and prediction accuracy, followed by Gradient Boosting Regression. However, in case of crops with smaller datasets, such as spring wheat, rapeseed and peas, linear regression provided better accuracy than more robust machine learning methods. The study emphasizes the importance of using the tandem of machine learning and remote sensing in modern agriculture to improve crop monitoring and yield prediction, contributing to sustainable agricultural practices and food security.