基于多源遥感影像与分层分类方法的水稻分布信息提取

Extraction of Rice Distribution Information Based on Multi-source Remote Sensing Image and Hierarchical Classification Methods

摘要:

水稻是全球最重要的粮食作物之一，准确高效地获取其种植分布信息对于保障农业生产、优化灌溉管理具有重要意义。本研究以江西省赣抚平原灌区为研究区，基于Google Earth Engine（GEE）云平台，构建了一种基于多源遥感影像与多特征参量的分层分类方法，在线获取并处理时序遥感影像，并进行分类计算，实现了2018---2022年研究区水稻种植分布信息提取。该方法第1层采用多种机器学习算法作为分类器进行土地利用分类，提取耕地信息。第2层在准确提取耕地分布基础上，结合作物物候特征，采用决策树分类算法实现水稻识别。结果表明，第1层土地利用分类中，随机森林算法表现最佳，精度最高;第2层水稻种植信息提取平均总体精度和Kappa系数达到93.29%和0.90，整体具有较好鲁棒性与较高提取精度。基于GEE云平台构建的分层分类方法不仅考虑了光谱、纹理特征等，还进一步考虑了作物物候特征并构建了多参量变化特征，既减少了对遥感影像质量的依赖，又保证了较高分类精度，同时适用于多年作物种植信息监测，体现较高计算效率、鲁棒性和可操作性。研究结果为实现遥感反演高精度的灌区作物种植分布信息提供了新思路和方法参考。

Abstract:

Rice is one of the most important food crops globally, and efficiently acquiring accurate information on rice planting distribution is crucial for ensuring agricultural production and optimizing irrigation management. Focusing on the Ganfu Plain Irrigation District in Jiangxi Province, based on the Google Earth Engine (GEE) cloud platform, a hierarchical classification method was developed by using multi-source remote sensing images and multiple feature parameters. Time-series remote sensing images were obtained and processed online, with classification calculations performed directly on the GEE. This process enabled the extraction of rice planting distribution from 2018 to 2022. In the first layer, multiple machine learning algorithms were employed as classifiers for land use classification to accurately identify cultivated land. In the second layer, building upon the extracted cultivated land, rice planting areas were further identified by using a decision tree algorithm integrated with rice-specific phenological characteristics. Results indicated that the random forest algorithm achieved the highest accuracy in the first layer of land use classification. For the second layer of rice mapping, the average overall accuracy (OA) and Kappa coefficient reached 93.29% and 0.90, respectively, demonstrating strong robustness and high precision. The proposed GEE-based hierarchical classification method effectively integrated spectral and texture features with crop phenological characteristics and constructed multiple feature indicators. This reduced dependence on the quality of remote sensing images and ensured high classification accuracy. Meanwhile, the method was applicable to multi-year monitoring of crop planting information, demonstrating high computational efficiency, robustness and operational feasibility. The proposed method can provide a reliable methodological reference for achieving high-precision remote sensing inversion of crop planting distribution in irrigated areas.