基于时空融合NDVI及物候特征的江汉平原水稻种植区提取研究

江汉平原是中国重要的商品粮基地,高精度的水稻种植面积的获取对国家的农业发展与规划具有重要意义。但是我国南方区域云雨天气较多,光学遥感影像缺失严重,同时受卫星重访周期的影响,可用数据较少,进而影响水稻种植面积提取的精度。为解决高时空分辨率影像缺失问题,基于ESTARFM (enhanced spatial and temporal adaptive reflectance fusion model)模型开展Landsat 8 OLI与MODIS数据的融合研究,获取具有高时空分辨率的Landsat NDVI时序数据。利用时序数据分析水稻的物候特征并结合关键物候期参数,采用多种机器学习方法对水稻种植区域进行提取。结果表明:利用该种方法能较好地提取研究区水稻种植的面积,并且在采用SVM方法分类时效果最好,水稻种植区域提取的总体分类精度为93.31%,Kappa系数为0.920 2。该研究为多云雨地区农作物种植信息提取提供了一种有效的方法。

Extraction of Rice Planting Areas in Jianghan Plain Based on Spatiotemporal Fusion NDVI and Phenological Characteristics

Jianghan Plain is an important commodity grain base in China, the acquisition of high-precision rice planting area is of great significance to the country’s agricultural development and planning. However, there are many cloud and rain weathers in southern China affected by climate, and the optical remote sensing images are seriously missing. At the same time, due to the satellite revisiting cycle, the available data is less, which affects the accuracy of rice planting area extraction. Obtaining high spatial-temporal resolution remote sensing images is the key to extracting rice growing areas in southern China. In order to solve the problem of high spatial-temporal resolution image loss, the fusion of Landsat 8 OLI and MODIS data based on Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) is used to obtain Landsat Normalized Difference Vegetation Index(NDVI) time series data with high spatial and temporal resolution. The ESTARFM model could improves the accuracy of heterogeneous landscape extraction for more heterogeneous and fragmented areas with high precision. In the existing crop area information extraction research, the classification relies on a single NDVI data, and the phenological feature information in the process of crop growth has not been fully utilized in the remote sensing classification structure. In this article, we use time NDVI series data to analyze the phenological characteristics of rice and combining key phenological parameters, so that a variety of machine learning methods could be used to extract rice planting areas, in this article, machine learning classification methods including Support Vector Machine(SVM), random forest and neural network were used to extract rice planting area and evaluate which method works best. The results show that this method can extract the rice planting area in the study area well, and the SVM method has the best classification effect. In the meanwhile, the overall classification accuracy of rice planting area extraction is 93.31%, and the Kappa coefficient is 0.920 2. This study provides an effective technical means for the extraction of rice planting area in the southern region, and providing technical support for regional land use planning and food policy.