Mapping intra-annual integrated crop-livestock systems in Mato Grosso using deep learning and MODIS time series

The Brazilian Amazon state of Mato Grosso has undergone rapid agricultural expansion and intensification in recent decades, leading to land depletion and high carbon emissions. Since 2010, the low-carbon agricultural Plan (ABC Plan) promotes practices to limit greenhouse gas emissions while increasing productivity. Among these practices, the adoption of intra-annual integrated crop-livestock systems (ICLS) is encouraged and can be monitored by remote sensing, especially through dense satellite image time series (SITS) and advanced analysis tools. In this study, we compare a calibrated Random Forest and three deep learning methods (LSTM, LTAE and TempCNN) suitable for processing SITS for mapping intra-annual ICLS in Mato Grosso. In particular, MODIS time series were used to classify a spatially and temporally rich dataset among four classes (pasture, single-crop, double-crop and intra-annual ICLS). TempCNN achieved slightly higher overall results in terms of precision (85.63%), recall (85.71%) and F1-score (85.32%), even though there is an inter-class and intraclass heterogeneity in the results. For the ICLS class, TempCNN achieved the best precision (85.14%) but Random Forest achieved higher recall (69.63% vs. 58.52% for TempCNN) and F1-Score (70.14% vs. 69.36% for TempCNN).