Large-scale crop type and crop area mapping across Brazil using synthetic aperture radar and optical imagery

• Published in: International journal of applied earth observation and geoinformation 2021-05, Vol.97, p.102294, Article 102294
• Main Authors: Ajadi, Olaniyi A., Barr, Jeremiah, Liang, Sang-Zi, Ferreira, Rogerio, Kumpatla, Siva P., Patel, Rinkal, Swatantran, Anu
• Format: Article
• Language: English
• Subjects: Crop classification; Deep learning; Harmonic function; Machine learning; Neural networks; SAR; Synthetic aperture radar; Time series; Xgboost
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2020.102294

•Easily updatable national-scale cropland mask using harmonic regression.•Easily updatable national-scale field boundaries using supervised deep learning.•Integration of generated features from Landsat, MODIS, and Sentinel-1 images from different geographies to predict crop type.•Evaluating crop classification models for its spatial and temporal transferability. Improved data on crop type and crop area from satellite imagery are invaluable for agronomy managers and are crucial for balancing agricultural expansion and forest degradation. However, large-scale maps of crop type and crop area using satellite imagery are not easily available in some regions, especially Brazil. Reasons for this include limited ground truth data, inadequate spatial and temporal satellite data availability, computational challenges, lack of cropland data and field boundaries. In this paper, we attempted to overcome some of these obstacles by using an ensemble of approaches to generate crop classification maps for Brazil. In order to compensate for the lack of abundant ground truth data in Brazil, we combined extensive field data and satellite input features from the United States with available field data and satellite input features from Brazil to train crop classification model for Brazil. Before applying the crop classification model for Brazil, we classified cropland areas using harmonic functions and delineated field boundaries using a supervised deep learning approach. Cropland masking and field boundary delineation allowed field-level mapping of crop type and crop area. Applying the crop classification model for Brazil in the states of Mato Grosso and Goias gave a true positive accuracy of 88% in the 2017/2018 summer growing season for soybean classification, 95% in the 2018 safrinha growing season for corn classification, and 86% in the 2018/2019 summer growing season for soybean classification. Our crop area estimates also showed a good agreement (correlation of 0.95 and mean absolute error of 0.64) with state-scale statistical data provided by the Companhia Nacional de Abastecimento (CONAB) in both summer and safrinha growing seasons adding further confidence to the results. These results suggest that extensive data from one geography can be used to train machine learning models in conjunction with limited field data from another geography. Accuracy assessments support the portability of crop classification model for Brazil with reasonable accuracy spatially, as tested i