关键词: 喀斯特山区, 石漠化耕地, 深度学习, 耕地制图, 耕地提取

Abstract:

Determining the accurate distribution of cultivated land is a prerequisite and foundation for the development of precise, modernized agricultural practices, and it is an important factor in land policies and agricultural production at the local scale. However, in karst areas, because of the complex terrain and cloudy and rainy weather conditions, the images often show repeated patterns of “identical foreign matter”; i.e., it is difficult to obtain accurate farmland information using traditional image-based farmland extraction. To solve this problem, Beipanjiang Town and Huajiang Town in Guizhou Province were selected as target areas in this study. Using Google’s high-resolution remote sensing images and combining the concepts of zoning and grading, accurate extraction and evaluation of cultivated land in the study area could be performed based on deep learning methods and traditional constraints. Firstly, based on visual data, farmland was divided into three types: farmland in gentle slopes), Relatively slender and unevenly distributed farmland, and farmland with blurred edges. Then, the Holistically-Nested Edge Detection (HED) model, Richer Convolutional Features (RCF) network model, and D_LinkNet semantic segmentation model were used to extract cultivated land. After confirming that the prediction data were sufficiently accurate, cultivated land was further classified based on the slope and topographical data of the test area, and the distribution characteristics of cultivated land within various terrains were studied. The results showed that: 1) The area of cultivated land in the study area classified via visual morphological differences and various models was 9 867 ha, which is basically consistent with the actual area. The F-measure was mainly distributed between [0.82, 0.98] and was affected by the topography and the exposure rate of rocks. The precision of arable land extraction in areas with severe rocky desertification was low. 2) In terms of morphological accuracy, the Generalized Intersection Over Union (GIOU) of the predicted and actual area of cultivated land was mainly distributed between [0.7, 1], The correct segmentation rate was more than 85%, which indicates that the predicted farmland boundaries generally coincided with the actual plot boundaries, and the extraction results reflected the actual situation in the study area. The research area mainly consists of karst cultivated land, which accounted for 74% of the total arable land area in the study area. Further, the degree of rocky desertification is relatively serious, of which light and moderate rocky desertified arable land accounts for about 32% of the total cultivated land. The narrowness and fragmentation of cultivated land in rocky desertified areas are greatly affected by anthropogenic activities. Land closer to residential areas is more accessible, and the tendency of the plots to be fragmented increases as the land utilization rate increases. In summary, the method proposed in this study can effectively extract cultivated land in areas with fragmented cultivated land and complex terrain, thus providing accurate data to support regional development, cultivated land management and research, and environmental protection and decision-making.

Key words: karst mountainous area, rocky desertified cultivated land, deep learning, cultivated land mapping, cultivated land extraction