Seasonal behavior and yield penalty classification in pasture areas based on SAR polarimetry and machine learning techniques

Pastures occupy extensive areas worldwide, especially in Brazil, and their use for cattle ranching promotes the extensive livestock competitiveness as a cheaper system. However, concerns about its sustainability emerge due to the degradation process present in most pasture areas, and so feasible monitoring methodologies to improve sustainability policies are demanded. Remote sensing solutions aiming at overcoming this problem concentrate efforts on optical sensors but still struggle to explore the Synthetic Aperture Radar (SAR) technology, which has fewer related studies and generally presents lower performance compared to traditional optical approaches. Here, we propose a model to classify pasture areas, previously characterized by seasonal patterns and yield penalization, through machine learning models built upon SAR polarimetric data. First, we used weather data from the Power (NASA) database and NDVI time series from MODIS-Terra to model pasture potential and observed yields and classified them based on seasonal patterns. Then we built predictive models based on SAR polarimetric parameters extracted from Sentinel-1, mainly the Stokes Vector, a coherent product that remains underused in SAR-based applications for pasture monitoring and compared the results achieved to a reference signal decomposition model. Our results reached an accuracy of up to 94% in pasture areas discrimination using the Random Forest algorithm in areas with a yield penalty ranging from ∼30% to ∼60%. So, our study aggregates the literature employing SAR for pasture monitoring using polarimetric data and machine learning techniques.