In-season winter cropland mapping with Sentinel-1

The issue of cloud coverage over agricultural areas impedes the use of optical Earth observation data for consistent in-season monitoring of cropland including area estimations. Sentinel-1 SAR has been proven to be a critical dataset in both complementing optical data and on its own in mapping crop-types and thus can overcome the issue of cloud coverage. However, there is still a lack of research on the efficacy of Sentinel-1 SAR for mapping winter cropland in major wheat producing countries such as Russia, and in particular at varying stages of the winter crop growing season. This gap highlights the need to provide research on how SAR data can address the challenge of cloud coverage that often obscures optical data, thereby improving timely and accurate cropland area estimations. In this study, we assessed the performance of mapping winter cropland at successive 15-day intervals, starting in the autumn planting period through to the end of the harvest period, using solely Sentinel-1 SAR VV (vertical-vertical) and VH (vertical-horizontal) bands in both ascending and descending orbits. Three models were utilized being a multi-layer perceptron model, a long short-term memory model, and a random forest model and their efficacy was assessed. Performance was assessed at each in-season time period in terms of user (UA) and producer accuracy (PA) in addition to area estimation comparisons with final year recorded winter cropland statistics at the sub-national level. Results indicate that in mid-November (planting year) (DOY -40), the UA was 0.68 and PA was 0.47; in early spring (DOY 100), the UA was 0.70 and PA was 0.78; and by early August, the UA reached 0.84 and PA 0.93. Notably, the LSTM model achieved even better results than the MLP model for estimating areas in the late autumn season (DOY -40) when comparing raw area differences (pixel count) from final reported statistics.