Mapping Crop Types across Mixed Single- and Double-Cropping Systems in Brazil Using Satellite Time-Series Data and Machine Learning
Shoki Shimada, and Kei Oyoshi

Japan Aerospace Exploration Agency (JAXA)

Abstract

Brazil, a major global agricultural exporter, plays a key role in the food supply chain, making accurate monitoring of its crop production essential for food security. Due to its vast area, satellite remote sensing is crucial for timely and cost-effective crop mapping, particularly given Brazil^s mix of double and single cropping systems, such as soybean-corn or soybean-cotton rotations and long-cycle crops like sugarcane. This study maps five key crops, soybean, corn, cotton, sugarcane, and grains (sorghum and millet), using Geen Cholophill Vgetation Index (GCVI) time series data from Landsat and Sentinel-2 between September 2023 and August 2024. The data were smoothed using the Harmonic ANalysis of Time Series (HANTS) algorithm in Google Earth Engine, and crop cycles were segmented through peak detection. Skewed normal distributions were fitted to each season^s phenology, and their parameters were used as input features for a Random Forest model to classify crop types in central Brazil. The model achieved an overall accuracy of 0.826, and municipality-level estimates for soybean, corn, cotton, and sugarcane showed strong agreement with official statistics, with R-squared values of 0.95, 0.95, 0.98, and 0.81 respectively. These results demonstrate the method^s effectiveness for accurate and timely crop mapping in regions with complex agricultural practices.

Keywords: Phenology, Random Forest, Food Security, Remote Sensing

Topic: Topic B: Applications of Remote Sensing