Evaluation of Deep Learning Approaches for Classifying Regional Croplands With Satellite Image Time Series in Senegal

With the significantincrease in the availability of 2-3 daily Harmonized Landsat and Sentinel-2 imagery (HLS), this satellite image time series (SITS) can capture the temporal and spatial features for accurate cropland mapping. This study evaluated the performance of four SITS deep learning (DL) classification approaches, including, convolutional LSTM, convolutional GRU, and 3D UNet, alongside a novel DPC-UNet architecture, for regional cropland mapping inSenegal, using 30 m HLS data. The study focused on the Eastern Casamance (ECAS) and Eastern Transition Zone (ETZ) regions. The study created extensive training data from multidate, 2 m WorldView imagery (WV) covering 15 locations (1500 km$^{2}$) and independently generated${>}$2000 validation points from the WV images. All the SITS DL models generated cropland maps with high overall accuracy${>}$92% in ECAS and${>}$80% in ETZ. All models displayed relatively low precision of 76 to 80% that was caused by the well-documented overestimation of croplands due to confusion with fallow or abandoned fields and bare sand along drainage lines. The relative performance of the models differed between regions, however, the ConvLSTM marginally had the highest F1 score and overall accuracy in both regions. DPC-UNet displayed the most overestimation in ECAS, but the least in the ETZ. The cropland maps produced by the DL models otherwise displayed strong agreement with the benchmark Global Land Analysis and Discovery (GLAD) cropland map and may also be detecting new expanding fields along the edges of the core cropland clusters. Given their performance and rate of progress, it is likely that SITS DL models will become the standard for operational cropland mapping in the near future, especially with the eminent, dense frequency, 10m HLS data.