POLARIMETRIC SAR DATA FROM SENTINEL-1A APPLIED TO EARLY CROP CLASSIFICATION
Keywords: Agriculture monitoring, Remote Sensing, Microwave, Soybean, Early classification, Machine learning
Abstract. This paper aims to map crops in two Brazilian municipalities, Luís Eduardo Magalhães (LEM) and Campo Verde, using dual-polarimetric Sentinel-1A images. The specific objectives were: (1) to evaluate the accuracy gain in the crop classification using Sentinel-1A multitemporal data backscatter coefficients and ratio (σ0VH, σ0VV and, σ0VH/σ0VV, denominate BS group) in comparison to the addition of polarimetric attributes (σ0VH, σ0VV, σ0VH/σ0VV, H, and α, denominate BP group) and; (2) to assess the accuracy gain in the earliest crop classification, creating new scenarios with the addition of the new SAR data together with the previous images for each date and group (BS and BP) during the crop development. For BS and BP groups, 13 e 10 scenarios were analyzed in LEM and Campo Verde, respectively. For the classification process, we used the Random Forest (RF) algorithm. In the LEM site, the best results for BS and BP groups were equivalent (overall accuracy: ∼82%), while for the Campo Verde site, the classification accuracy for the BP group (overall accuracy: ∼80%) was 2% higher than the BS group. The addition of new images during the crop development period increased the earliest crop classification overall accuracy, stabilizing from mid-February in LEM and mid-December in Campo Verde, after 10 and 8 images, respectively. After these periods, the gain in classification accuracy was small with the addition of new images. In general, our results suggest the backscattering coefficients and polarimetric attributes extracted from the Sentinel-1A imagery exhibited a great performance to discriminate croplands.