Assessing SAR C-band data to effectively distinguish modified land uses in a heavily disturbed Amazon forest

Abstract The Amazon is the largest expanse of tropical rainforest globally and deforestation resulting from land use changes poses a major concern for sustainable resource management. Synthetic aperture radar (SAR) data have all-weather and all-day capability, and thus are well-suited for mapping land use land cover (LULC) in tropical regions, which are seasonally influenced by cloud cover. Understanding modified land uses and drivers of deforestation is fundamental for the development of policies and measures to reduce emissions and for developing forest reference levels. Sentinel-1 C-band SAR data present unprecedented potential since the observations are free and openly available, providing for the first regular and standardize SAR data. This study analyzes the applicability of Sentinel-1 data for LULC classification as an effort to differentiate modified land uses, which is a current need for early-warning deforestation systems. The study area covers a deforestation frontier in the Peruvian Amazon where the landscape is characterized by a mosaic of LULC types. Collect Earth Online is used for reference LULC data collection, and seven classes are defined for this study: forest, secondary vegetation, agriculture, pasture, urban, mining, water. Amplitude γo time-series spanning 2017–2019 are analyzed along with statistical metrics for each class, and a classification decision tree is developed in Google Earth Engine. Overall accuracy obtained is considered low (52%). Results show high user's accuracy for forest and water classification, a lot of confusion between agriculture, secondary vegetation, and forest, and the use of the polarization ratio VV/VH is suggested to be useful for pasture classification. The orientation of streets in a urban environment is confirmed to have high influence on backscattering response. This study provides information for future research on LULC and the identification of drivers in deforestation monitoring systems that could result in additional actionable information for decision-making.