QIN Xiaowei, CHENG Bo, YANG Zhiping, LI Lin, DONG Wen, ZHANG Xin, YANG Shuwen, JIN Zongyi, XUE Qing
The distribution of crop planting structure can directly reflect the form and location of farmland, which is of great significance for the development of precision agriculture and inventory of agricultural resources. In the southwestern mountainous area in China, the farmland parcel is small and fragmented, with complex planting types, and thus the image visual characteristics are fuzzy and diverse. Due to the complexity of climate and terrain surface, it is difficult to identify crop types in this area. The arrival of the era of high-resolution remote sensing has promoted the rapid development of object-oriented crop extraction methods. Classification based on object segmentation combing spectral, contextual, and other features can effectively overcome the shortcomings of pixel-based classification methods and better serve the field farming management and land ownership management. Therefore, this study uses high-resolution remote sensing images to extract farmland parcels. Taking parcels as the spatial constraints, six time-series vegetation indices including NDVI, EVI, RVI, DVI, GNDVI, and SAVI, are constructed at the scale of parcels based on multi-temporal medium-resolution Sentinel-2 and Landsat8 OLI remote sensing images. Their phenological characteristics are also extracted using these remote sensing images. The multi-dimensional data are constructed through the importance evaluation of classification features, and the random forest classification model is used to identify crop types (i.e., lemon, rape, rice, corn, and other crops) combined with field sampling data. By comparing the classification accuracy using vegetation indices or phenological characteristics alone and that combining vegetation indices and phenological characteristics together, we found that the classification accuracy combining vegetation indices and phenological characteristics is the highest. The overall accuracy is 94.52%, and the Kappa coefficient is 0.90. The accuracy of lemon, rape, corn, rice, and others is 89.88%, 87.30%, 84.98%, 95.57%, and 98.05%, respectively. There are 90 348 lemon parcels, 30 173 rape parcels, 271 606 corn parcels, 146 113 rice parcels, and 16 406 other parcels, accounting for 10.96%, 36.40%, 32.94%, 17.72%, and 1.99% of the total farmland in Tongnan area, respectively. The planting area accounted for 14.82%, 18.44%, 35.37%, 29.92% and 1.45%, respectively. Our results indicate that the vegetation indices and phenological characteristics extracted from time-series remote sensing data can be used for crop identification at the parcel scale, which provides reference for obtaining large-scale crop mapping in the region with complex planting structures.