Abstract

Soybean mapping in Indian context is challenging owing to its short growing period coinciding with the monsoon clouds, inter-cropping and smallholders’ land. This study proposes an approach for mapping soybean by integrating object-based image analysis with machine learning (ML) based classification using temporal Sentinel-1 SAR (S-1) and Sentinel-2 optical (S-2) data. Field objects were delineated with scale-optimized multi-resolution segmentation using historical S-2 data. Object-based temporal VH-backscatter and NDVI were extracted for training, validation and testing of the three ML models. Validation results showed the outperformance of Extreme gradient boosting (OB-XGBoost) over Random Forest (OB-RF) and Support vector machine (OB-SVM) with an overall accuracy (OA) of 92.50, 91.08 and 90.1, respectively. Testing of OB-XGBoost model resulted in OA, kappa statistics, and F-score (soybean) of 86.12%, 0.82, and 87.23%, respectively. The soybean map produced by the proposed methodology has shown better representation in terms of homogeneity and uniformity than the pixel-based classification.

摘要

印度的大豆制图具有挑战性，因为它的生长期短，与季风云、间作和小农土地重合。本研究提出了一种通过使用时间 Sentinel-1 SAR (S-1) 和 Sentinel-2 光学 (S-2) 数据将基于对象的图像分析与基于机器学习 (ML) 的分类相结合的大豆制图方法。使用历史 S-2 数据通过比例优化的多分辨率分割来描绘现场对象。提取了基于对象的时间 VH 反向散射和 NDVI，用于三个 ML 模型的训练、验证和测试。验证结果表明，极限梯度提升 (OB-XGBoost) 优于随机森林 (OB-RF) 和支持向量机 (OB-SVM)，总体准确度 (OA) 分别为 92.50、91.08 和 90.1。OB-XGBoost 模型的测试导致 OA、kappa 统计量和 F 分数（大豆）分别为 86.12%、0.82 和 87.23%。与基于像素的分类相比，所提出的方法生成的大豆图在同质性和均匀性方面表现出更好的表现。