Mapping pasture management in the Brazilian Amazon from dense Landsat time series

Abstract

Intensification of cattle ranching has the potential to reduce deforestation rates in the Brazilian Amazon by decreasing the demand for new agricultural land. Explicit spatial knowledge on where, when and how pastures are managed and intensification takes place is needed to better estimate potentials of more sustainable management. Monitoring the frequency of management practices like burning of pasture land and tillage treatment with adequate spatial resolution therefore offers novel indicators for describing land use intensity. With dense time series of Landsat data, it appears possible to quantify land use intensity also in heterogeneous landscapes where fine-scale processes cannot be monitored with previously available datasets.Our overarching goal is to describe the occurrence or absence of extensive or intensive management regimes over time. For this study, we focused on detecting fire and tillage events in the region of Novo Progresso, Pará, Brazil, where deforested land is mostly used for cattle ranching by both largeholders and smallholders. We used a dense time series of Landsat-7 and Landsat-8 surface reflectance data to mitigate the problem of varying cloud cover. For each acquisition date, we extracted a temporal sequence of three subsequent clear observations at pixel level. The temporal variation in each clear observation sequence was characterized by a stack of spectral and temporal features. These feature stacks were classified with a random forest to identify the management events. We aggregated the classification results based on the random forest class probabilities and derived normalized annual class scores for land management events. The yielded class score maps show burned and tilled areas on a yearly basis and provide a measure of model confidence.We detected burned pastures with area adjusted user accuracies of 80–98%, burned secondary regrowth with 63–80% and tilled pastures with 74–78%. Our approach was able to discriminate management events even when they succeeded very fast. This way, the mapping approach with clear observation sequences allows us to contextualize management events directly from dense time series of high resolution satellite data. This opens new pathways to disentangle how management practices between smallholders and agribusinesses vary in the Brazilian Amazon. With sensor constellations from Sentinel-2 and Landsat data becoming a unified source for much denser time series soon, our method bears great potential to better understand and map pasture dynamics in the Amazon.