|
The North China Plain is one of the major grain producing areas in China, and most of the farmland in the grain growing areas has been in the state of "fragmentation and heterogeneity" for a long time, which makes it more difficult to recognize and extract typical crop features in the remotely sensed images. This study takes Gaofen-6 remote sensing images as the research object, takes Beijing-Tianjin-Hebei as the research area, combines the characteristics of farmland landscape zoning, and utilizes artificial intelligence algorithms to realize the remote sensing intelligent interpretation of winter wheat. The remote sensing identification of crops is affected by the distribution of crops, and this paper screens four indices that can characterize the spatial distribution of crops from the spatial structure of the landscape, spatial characteristics, and landscape diversity indices. Combined with the k-mean mean for cluster analysis, the study area was divided into four different farmland type zones. Random Forest (RF), U-Net and Deformable Full Convolutional Neural Network (DFCN) were utilized to extract winter wheat from the four farmland type areas, determine the optimal decoding method for each farmland type area, and finally combine the optimal models of the four farmland type areas to achieve the optimal decoding method. The optimal models of the four types of farmland areas are combined to realize the global optimal decoding of the study area. The research results show that: (1) the quantitative description index of crop distribution can be realized by four landscape indices, namely: shape factor LSI, area-weighted average patch dimension FRAC_AM, agglomeration index AI and uniformity index SHEI; through the k-mean clustering classification method, all landscape units can be clustered and analyzed, which can be classified into four typical farmland zoning types. types of farmland zoning. (2) DFCN, U-Net and RF were used to decipher winter wheat in each farmland type area. According to the comparison of pixel accuracy, the optimal model for winter wheat recognition in the first, third and fourth types of areas was DFCN, with an accuracy of 95.8%, 94.1% and 95.3%, respectively; and the optimal model for winter wheat recognition in the second type of areas was RF, with an accuracy of 93.6%. The idea of partitioning before decoding in this paper improves the decoding accuracy of winter wheat compared with the remote sensing decoding of winter wheat with the same method for the whole study area, and provides reference value for crop classification under the complex terrain conditions.
|
