Supervised machine learning in crop recognition through remote sensing: A case study for Ukrainian croplands

Abstract

Pavlo Lykhovyd, Raisa Vozhehova, Iryna Bidnyna, Oleksandr Shablia, Oleksandr Averchev, Nataliia Avercheva, Valerii Kozyriev, Tetiana Marchenko, Liubov Leliavska, Oleksandr Haydash, Maiia Hetman, Olena Piliarska

Automated crop recognition is an important branch of modern agriculture. It provides wide opportunities for cropland mapping, crop rotations analysis, cropland structure and agricultural land use monitoring, etc. Remote sensing is a prospective and powerful technique for crop recognition through the implementation of various vegetation indices, e.g., normalised difference vegetation index, in combination with technologies of machine learning and computer vision. Current study is devoted to real-world testing of the accuracy of recent development in supervised machine learning for crop recognition in Ukraine, namely, software application Agroland Classifier, which has been built based on the results of scientific research at the Institute of Climate Smart Agriculture of NAAS. The application utilizes several supervised machine learning approaches, namely, multiple canonical discriminant analysis and logistic regression, to distinguish between such crops as winter wheat, winter barley, winter rapeseed, grain maize, soybeans, and sunflower. The testing was carried out using randomly chosen labelled fields with known cultivated crops, 100 fields per each crop. Testing was carried out throughout all the territory of Ukraine. The input values of monthly normalised difference vegetation index were retrieved from Agromonitoring Crop Map platform. It was established that the highest precision of crop recognition was associated with wheat (overall accuracy of 82.0%, F1 score 0.90), while the worst results were recorded for soybeans (50.0% of true guesses, F1 score 0.67). It was also observed that the recognition accuracy is highly dependent on soil-climate conditions of the crops cultivation. Further detailed testing and algorithms improvement are required and will be held on.