首页 > , Vol. , Issue () : -
高分辨率遥感影像中耕地特征复杂，人工目视解译和传统的遥感影像分类方法提取能力有限，无法实现大范围的自动化高精度耕地提取。深度学习技术因具有较强的地物表达能力，在遥感影像信息自动提取方面表现出了优越的性能，为大范围耕地的精细化自动提取提供了新的思路。探究不同典型网络模型在不同景观特征耕地提取上的适用情况对耕地提取质量和效率的提升具有重要意义。基于此，本研究以高分一号及高分二号融合的2米分辨率数据为数据源，采用改进的金字塔场景解析网络(Modified Pyramid Scene Parsing Network，MPSPNet)和UNet网络模型，应用于山东省的耕地精细自动化提取，并与传统面向对象的方法对比，探究两种深度卷积神经网络模型在大尺度耕地自动提取中的适用性。研究获得以下结论：(1) MPSPNet模型和UNet模型在区/县尺度的耕地提取上性能优于传统的面向对象的分类方法，在全省尺度的耕地提取上总体精度优于90%且无明显差异。(2) 耕地景观特征是影响两模型耕地提取效果的重要因素，模型的选择对耕地提取效果无明显影响。在耕地景观指数较低的地块规则平整的区域，模型提取效果较好，在耕地景观指数较高的地块破碎丘陵区域以及与耕地特征相近的地块区域，模型提取效果较差，并且UNet模型在这些区域误分耕地的概率更大。(3) 两模型在不同区域、不同时相的影像中能得到较好的耕地提取效果，具有较强的泛化能力和时空迁移能力。
The rapid development of remote sensing image technology enables a large number of high-resolution remote sensing images to provide good data support for the accurate extraction of cropland and other ground features. However, high-resolution remote sensing images have large data volume and complex features, the artificial visual interpretation and traditional classification methods have limited extraction capabilities which cannot realized large-scale high-precision cropland extraction automatically. Deep learning technology has shown superior performance in the automatic extraction of remote sensing image information due to its strong ability to express features, providing a new idea for the automatic extraction of large-scale cropland. Exploring the application of different typical network models in the extraction of cropland with different landscape features is of great significance to the improvement of the quality and efficiency of cropland extraction. Based on above, the study uses the 2m resolution data fused with GF-1 and GF-2 in 2015-2017 as the data source. Using Modified Pyramid Scene Parsing Network (MPSPNet) and UNet models applied to the fine automatic extraction of cropland in Shandong Province, and compared with the traditional object-oriented method, exploring the applicability of two deep convolutional neural network models in the automatic extraction of large-scale cropland. We also apply the trained models to the images of different regions and different time phases for the extraction of cropland, and explore the generalization ability of the models. The landscape features of cropland and uncertainty results are analyzed to explore the factors affecting the accuracy of cropland extraction by the models. Results show that: (1) MPSPNet and UNet models perform better than traditional object-oriented classification methods in the extraction of cropland at the district/county scale, the overall accuracy of the extraction of cropland at the provincial scale is better than 90% and there is no obvious difference between two models. (2) The landscape characteristic of cropland is an important factor that affects the effect of the two models, and the choice of the model has no obvious influence on the cropland extraction effect. The extraction effect is better in areas where the cropland landscape index is low and the plots are regular and flat, and the extraction effect is poor in the broken hilly areas of the plots with high cropland landscape index and in the noncropland plots whose characteristics are similar to the cropland, the UNet model is more likely to misclassify cropland in these areas. (3) The two models can obtain better cropland extraction effects in images of different regions and different time phases, and have strong generalization capabilities and temporal and spatial migration capabilities. This study proves the powerful feature learning capabilities of MPSPNet and UNET network models for high-resolution images, and the application potential of deep learning algorithms in fully automatic high-resolution cropland extraction.