Abstract
Automatic cropland monitoring is becoming increasingly important in the advancement of sustainable agriculture. However, multiannual satellite-based crop mapping across different regions remains challenging due to variations in crop phenology and meteorological conditions. The use of multispectral data from a single satellite can also present difficulties in constructing vegetation index time series, particularly in regions affected by persistent cloud cover. In this study, NDVI time series obtained from Sentinel-2 and Landsat-8/9 imagery were fitted using a Fourier series, and daily NDVI composites from the Meteor-M satellite were obtained for Khabarovsk Krai in the Russian Far East from 2022 to 2024. These data were used to perform random forest (RF) classification for each year for five land cover classes: soybean, grain crops, perennial grasses, buckwheat, and fallow land. The average annual classification accuracies were 87% for Landsat 8/9, 89% for Meteor-M, and 93% for Sentinel-2. Combining data from all three satellites improved classification performance, increasing cross-validation overall accuracy from 92% to 96% in 2022, from 96% to 97% in 2023. These results demonstrate the potential of using both individual satellite data for sufficiently accurate mapping and combined datasets, which provide consistently high accuracy and are a reliable alternative when Sentinel data are limited due to cloud cover.