Investigation of the accuracy of ground reference datasets using multi-temporal Sentinel-2 satellite images: A case study with barley and wheat crops

Abstract

There have been many studies regarding agricultural crop classification using remotely sensed imagery in Turkiye. In such studies, crop reference datasets are needed to train classification models and validate the model results. However, some studies reported that there are some inaccuracies in these crop reference datasets. In this study, the accuracy of the crop reference dataset collected by a commercial company in the central district of Yozgat Province was investigated using the imagery acquired by Sentinel-2A and Sentinel-2B satellites. The Normalized Difference Vegetation Index (NDVI) time series of each agricultural parcel were constructed and iteratively compared to the average NDVI time series for a given crop. Later, statistic metrics such as correlation and average minimum distance were used to identify incorrectly labeled agricultural parcels. The method is fully automated and doesn't need any user intervention except for the steps such as the data download and the modification of parcel boundaries. Therefore, The automation of the method steps was carried out in Python programming language using various open-source Python libraries.