Enhancing Early-Season Soybean Identification through Optical and SAR Time-Series Integration

Hongchi Zhang^1,2Dailiang Peng^1,2*Changyong Dou^1,2*Zihang Lou³XIAOYANG ZHANG⁴Le Yu⁵Kaishan Song⁶Yaqiong Zhang⁷Jingkang Hu^1,2Shijun Zheng⁵Yulong Lv^1,2Shengyi Liu⁸Yizhou Zhang^1,2Hao Peng⁹

• ¹Key Laboratory of Digital Earth Science, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
• ²Chinese Academy of Sciences International Research Center of Big Data for Sustainable Development Goals, Beijing, China
• ³Zhejiang University College of Environmental and Resource Sciences, Hangzhou, China
• ⁴Geospatial Sciences Center of Excellence, Brookings, United States
• ⁵Tsinghua University Department of Earth System Science, Beijing, China
• ⁶Northeast Institute of Geography and Agroecology Chinese Academy of Sciences, Changchun, China
• ⁷Ministry of Ecology and Environment Satellite Application Center for Ecology and Environment, Beijing, China
• ⁸Shenyang Agricultural University College of Forestry, Shenyang, China
• ⁹Xinjiang Institute of Ecology and Geography, Urumqi, China

Soybean is an important grain and cash crop in China, and timely knowledge of its distribution is crucial for food security. However, traditional survey methods are time-consuming and limited in coverage. In contrast, satellite remote sensing enables large-scale, continuous, and cost-effective monitoring, providing reliable support for crop classification and yield forecasting. Yet the high spectral similarity between soybean and maize during key phenological stages presents a major challenge for reliable classification. To address this, we propose a multi-source remote sensing approach that integrates Sentinel-1 SAR and Sentinel-2 optical time-series imagery. This method combines statistical descriptors, harmonic fitting parameters, phenological indicators, and radar-based features within a random forest classifier to achieve accurate soybean mapping. The study was conducted in the Jiusan Reclamation Area of Heilongjiang Province using satellite imagery from May to October 2019 for multi-source classification and temporal analysis. We systematically evaluated classification performance across different data sources and phenological stages and introduced the Earliest Identifiable Time (EIT) metric to assess temporal detection capabilities. Results show that the multi-source fusion approach outperforms single-source methods, achieving an overall accuracy (OA) of 96.85%, a Kappa coefficient of 0.9493, and an F1-score of 95.84% for soybean. Notably, SAR data significantly improved classification during the flowering stage—when optical imagery is often constrained—resulting in a maximum F1-score increase of 6.96%. Soybean classification accuracy increased rapidly with crop development, and the EIT was advanced to Day of Year (DOY) 210, approximately 20 days earlier than with optical data alone. These findings demonstrate the effectiveness of multi-source remote sensing in enhancing both the accuracy and timeliness of crop classification under complex climatic conditions, offering valuable support for precise soybean mapping and in-season monitoring.