基于HLS遥感数据与面向对象分割的农作物分类研究

Crop Classification Using HLS Remote Sensing Data and Object-Based Image Analysis

摘要: 农作物分类是农业遥感研究的重要内容，对于农业资源管理和作物种植结构监测具有重要意义。为提高农作物分类精度，本文以黑龙江省七台河市茄子河区为研究区，基于Sentinel-2光学遥感数据和HLS (Harmonized Landsat Sentinel)遥感数据，结合SNIC面向对象分割方法与随机森林分类算法，对研究区主要农作物类型进行了识别与分类。首先对遥感影像进行云掩膜和预处理，构建多时相NDVI等遥感特征，并利用SNIC算法对影像进行超像素分割，提取对象尺度特征；随后基于随机森林模型进行农作物分类，并通过混淆矩阵对分类结果进行精度评价。结果表明：基于Sentinel-2光学遥感数据时，基于像素方法和面向对象方法的总体精度分别为88.86%和90.39%，Kappa系数分别为0.8659和0.88，面向对象方法在整体分类效果上略优于基于像素方法。在此基础上引入HLS多源遥感数据后，分类总体精度进一步提高至92.55%，Kappa系数达到0.91，其中水稻的识别精度提升最为明显。研究结果表明，多源遥感数据融合能够提供更加丰富的时间序列信息，而面向对象方法能够有效减少像素级噪声，两者结合能够显著提升农作物分类精度。

Abstract: Crop classification is an important topic in agricultural remote sensing and plays a significant role in agricultural resource management and crop planting structure monitoring. To improve classification accuracy, this study takes Qiezihe District of Qitaihe City, Heilongjiang Province, as the study area, and conducts crop classification based on Sentinel-2 optical remote sensing data and Harmonized Landsat Sentinel (HLS) data, combined with the SNIC object-based segmentation method and the random forest classification algorithm. First, cloud masking and preprocessing were applied to the remote sensing images, and multi-temporal features such as NDVI were constructed. Then, the SNIC algorithm was used to perform superpixel segmentation and extract object-level features. Subsequently, crop classification was carried out using a random forest model, and the classification results were evaluated using a confusion matrix. The results show that, based on Sentinel-2 data, the overall accuracies of pixel-based and object-based methods are 88.86% and 90.39%, with Kappa coefficients of 0.8659 and 0.88, respectively, indicating that the object-based method performs slightly better. After incorporating HLS multi-source data, the overall accuracy further increases to 92.55%, with a Kappa coefficient of 0.91, among which the classification accuracy of rice shows the most significant improvement. The results demonstrate that multi-source data fusion provides richer temporal information, while object-based methods effectively reduce pixel-level noise. Their combination can significantly improve crop classification accuracy.