地块尺度的复杂种植区作物遥感精细分类

实现复杂农区作物种植信息的精准、动态监测是中国农业精细化管理面临的迫切需求，而作物种植碎片化和异质性给作物遥感精细分类带来了诸多挑战，该文旨在探索基于高分辨率影像的地块尺度多种作物同步识别方法，以满足实时获取复杂农区作物详细分布信息需要。研究选取武汉市新洲北部为典型区，以WorldView-2影像为数据源，利用ReliefF-Pearson方法优选作物遥感特征，采用人工神经网络、K最近邻和随机森林算法进行作物分类，并对比分析其精度。研究发现：1）RVI、NDVI、相关性和边界长度等12个特征构成了地块尺度作物分类的相对较优特征，可在充分表征影像信息同时降低数据冗余；2）相比于人工神经网络和K最近邻算法，随机森林算法分类精度最高，其总体精度达79.07%；3）以光谱特征差异为作物区分基础，形状和纹理特征的使用能有效改善地块尺度作物分类精度，总体精度可提高4%左右；4）研究所采用的方法体系能有效提升复杂种植区地物分类精度，水稻、棉花、荷等主要作物以及裸旱地、裸水田等地物分类精度均达到了80%以上。研究成果可为复杂种植区作物遥感精细分类提供新的思路和方法借鉴，亦可为作物种植信息精准普查、土地利用精细化管理以及农业产业结构调整动态监测等提供参考。

Fine crop classification by remote sensing in complex planting areas based on field parcel

Timely and accurate information of crop planting structure is of great significance for monitoring agricultural conditions, estimating crop yield, adjusting agricultural structure and formulating food policies. However, currently only little explicit information about spatial crop patterns is known, especially in China where the farmland landscapes are extremely fragmented and heterogeneous. At present, techniques for quantifying crop spatial patterns may be insufficient to map crops in complex planting areas, the plot sizes of which are smaller than the spatial resolution of ready-to-use satellite data. In order to achieve the fine mapping of crops in complex planting areas, this study aimed to explore approaches that simultaneously mapping multiple crops on parcel scales with high spatial resolution images. A 2.9 km×2.6 km complex heterogeneous planting area in the suburb of Wuhan, Hubei Province was selected as the typical study area. Combined with high spatial resolution images, an improved method of fine crop mapping based on geo-parcels was presented. Using the spectral, shape and texture information of images, combined with random forest (RF), artificial neural network (ANN), and K-nearest neighbor (KNN) algorithms, WorldView-2 images were accurately classified through the following steps. First, Worldview-2 images were visually interpreted to obtain the distribution of cropland and non-cropland in this study area, so as to mask out non-cropland information in remote sensing images. Second, the pre-processed WorldView-2 images were segmented by using the land parcel boundary vector data obtained from manual visual interpretation, and 32 feature variables of the image object were extracted, including NDVI, area, GLCM-correlation, etc. Third, the ReliefF-Pearson feature dimension reduction method was adopted to remove redundant features with high correlation and weak classification ability. Then, RF classification was performed with optimal features and field sampling data, and the accuracy of the classification results was evaluated. Subsequently, the accuracy of RF classification was compared with that of ANN and KNN to verify the effectiveness of RF algorithm. Finally, four feature combination sets were constructed based on optimal features, and RF classification accuracy was compared under four feature combinations to evaluate the contribution of shape and texture features to crop classification. The results showed that 1) The 12 feature variables, such as RVI, NDVI, GLCM-correlation and border length, were the optimal features of parcel-level crop classification based on high spatial resolution images, which can fully characterize image features and reduce data redundancy; 2) The RF method had the highest classification accuracy, with an overall accuracy of 79.07%, kappa coefficient of 0.76, and the overall accuracy of KNN and ANN method was above 70%; 3) Compared with the method of only using spectral features, adding shape or texture information could effectively improve the accuracy of crop classification, and the overall accuracy could be improved by 3.86% and 3.05%, respectively; 4) Based on the optimal features and RF classification method, the classification accuracy of rice, cotton, lotus, bare upland field and bare paddy field was over 80%, while that of abandoned cropland, peanut and ''other crops'' was only about 60%. This study provides new ideas, methods and technical means for realizing the fine classification of crops by remote sensing in complex planting areas, and can provide references for accurate survey of crop planting information, refined management of rural land use and dynamic monitoring of agricultural industrial structure adjustment. In the future, the image segmentation technology will be further studied to ensure the consistency between segmentation objects and geographical entities as much as possible, and improve the accuracy and automaticity of crop remote sensing classification on parcel scales.