Satellite-based data fusion crop type classification and mapping in Rio Grande do Sul, Brazil.

• Data fusion offered new opportunities for agricultural monitoring. • Spatial data patterns affected model performance in regions with high variability. • Aggregating crop year data presented high accuracy in transfer learning. • Early crop classification resulted in similar overall accuracy to full length time-series. • Crop classification and mapping were produced for Rio Grande do Sul state, Brazil. Field-scale crop monitoring is essential for agricultural management and policy making for food security and sustainability. Automating crop classification process while elaborating a workflow is a key step for reliable and precise crop mapping. This study aims to develop an approach for crop classification in the state of Rio Grande do Sul, Brazil, following the specific goals of i) evaluating spatial satellite-based features to guide crop data collection; ii) testing transfer learning model with subsequent growing season data; iii) examining accuracy in early-season prediction model; and lastly, iv) developing a crop classification model for estimating large scale crop area. As main data inputs, Sentinel-2, Sentinel-1, and Shuttle Radar Topographic Mission (SRTM) Digital Elevation data were used to extract features to input in the Random Forest classifier. Spatial variability of satellite features was evaluated using Moran's I Index and cluster k-means. Crop area prediction data were obtained at municipality level to compare with census data (standard method). A crop summer map layer was generated for three major crops: soybeans (Glycine max L.), corn (Zea mays L.), and rice (Oryza sativa L.) in the state of Rio Grande do Sul, Brazil. The crop classification model achieved an overall accuracy of 0.95. Model performance was influenced by sample size and spatial variability of the samples. The random forest model was transferred to the next growing season with 0.89 and 0.91 overall accuracy for 250 and 750 samples, respectively. However, overall accuracy increased from 0.93 to 0.95 when 50 to 250 samples of same-year data was aggregated to the model. Similar accuracy was obtained for predictions done with data until March relative to when the entire season was considered, until May. When data for more growing seasons were aggregated, the model produced more accurate early season predictions (January and February). Soybean prediction area obtained the highest performance (R2 = 0.94), relative to rice (R2 = 0.90) and corn (R2 = 0.37). The rice prediction area presented a high precision, but the crop area was overestimated due to errors with wetland target relative to other class. Lastly, this study presents the first crop map layer of the three major field crops for the state of Rio Grande do Sul, Brazil, serving as a foundation for the creation of crop type maps for other states in the country and around the globe. [ABSTRACT FROM AUTHOR]