Using regional normalized difference vegetation index for the large-scale yield prediction of potato, vegetables, fruits, and berries, cultivated in Kherson region of Ukraine

Abstract

Pavlo Lykhovyd, Raisa Vozhehova, Oleksandr Averchev, Oleksandr Rudik, Liudmyla Hranovska, Sergiy Lavrenko, Nataliia Avercheva, Grigoriy Latiuk

Large-scale yield prediction of major crops is important to ensure rational science-based policy in agricultural economic activity, especially in planning import and export of plant products and food security assessment. Remote sensing is flexible and convenient tool for the evaluation and prediction of crop yields on large areas. In this study, remote sensing data on the normalized difference vegetation index, calculated for the croplands of Kherson region, were applied to create regression models of potato, vegetables, fruits, and berries yields in the region. The normalized difference vegetation index values were calculated using raw MODIS Terra images for the croplands of the region, retrieved from the service of the University of Natural Resources and Life Sciences (Vienna, Austria) and GIMMS Global Agricultural Monitoring System for the period 2005-2021. The average annual yields of the crops in the region were retrieved from the Ukrainian State Statistical Service for the stipulated period. As a result, linear regression models and models based on artificial neural networks were created to predict yields based on the values of the normalised difference vegetation index. The strongest relationship between the remote sensing data and yield was established for vegetables (in May, R=0.63), while the weakest relationship was established for fruits and berries (in August, R=0.33). The regression models developed have a reasonable to good prediction accuracy for potatoes and vegetables (MAPE=10.04—21.07%), while the prediction of fruits and berries yields has a low precision and reliability. The developed models could be further used in agrarian policy substantiation in Kherson region, as well as in scientific purposes. Artificial neural network-based models provided better predictive accuracy but are less helpful in understanding the principles of regional crop yield prediction.