关键词: 遥感提取, 作物种植结构, 归一化植被指数, 时间序列, 黑河流域中游地区

Abstract:

Timely and accurate acquisition of cropping structure is crucial for regional water resource management and crop yield estimation. Remote sensing of cropping structure based on multiple-temporal imagery can make full use of temporal featuresfor especially regions with complicated cropping structure.However, how to select remote sensing data and classifier remains a challenge.In this paper, to map the crop distribution and planting characteristicsin the middle reaches of the Heihe River Basin, we proposed a new method by combining multi-temporal remote sensing imagery and multiple classifiers. Eighteen images available from the Gaofen-1（GF-1） satellite in 2018 were applied to construct the time series of normalized differential vegetation index （NDVI） according to the following hierarchical principles: from easy to difficult, from specific to general. The level-1 and level-2 land cover categories were interpreted successively. Specifically, in interpreting level-1 land cover categories, the decision tree classification method was used to ide.pngywater bodies based on NDVI; and then, the object-oriented classification method was used to interpret construction land by NDVI texture information after zoning; and finally, the random forest classification method was used to classify cropland, forest land, grassland, bare land, and wetland. To further classify cropland, the decision tree classification method was used firstly to interpret alfalfa which has special phenology regularity and is easy to distinguish, and then to interpret wheat categories which possess large phenological differences from other categories, and finally to interpret corn, vegetables, and other crops that have similar phenological conditions. The overall accuracy of the level-1 land cover classification and cropping structure interpretation in the study area is 97.24% and 86.58%, respectively, with the kappa coefficient being 0.96 and 0.80, respectively. In addition, four factors affecting the interpretation accuracy of the research area in the middle reaches of the Heihe river basin are analyzed: definition of landcover categories, mixed pixels, selection of basic image in image segmentation and selection of classification method. By comparing different classification methods, Compared with methods which employ only a single classifier （e.g., maximum likelihood classification method, support vector machine classification method, and random forest classification）, the proposed method has higher interpretation accuracy.

Key words: extraction of remote sensing, cropping structure, normalized vegetation index, time series, the middle reaches of the Heihe river basin