Using Interannual Transfer Learning for Cropland Classification in the Khabarovsk District

Abstract

Automated crop mapping using multiannual remote sensing data is one of the main tasks in digital agriculture. Approach, using Normal Difference Vegetation Index (NDVI) time series and Random Forest (RF) classifier, was developed. Time series function-fitting using Fourier series was performed to align sample quantity and bring series for 2021 and 2022 to the same timeline. Approximated weekly frequency NDVI time series were used as input to the classifier. Time series labeled with one of the three classes (soybeans, oats, perennial grasses) for 2021 were used to train the classifier. The labeled NDVI time series for 2022 was used as a test set. The overall accuracy of interannual transfer learning was 88.5%. The F1 for soybean was 0.93, for oat - 0.68, and for perennial grasses - 0.53. This approach can be used for crop mapping in regions with the same crops, crop phenology and climatic conditions.