深度特征融合的黄土微地貌精细化提取模型研究

黄土微地貌表征了微小尺度下物质、能量随时间在地表的分配结果。针对黄土微地貌的提取能够为土壤流失及水土保持相关研究提供本底数据，同时也可作为研究地貌演化发育的基础性数据集。现阶段针对黄土微地貌的提取研究大多是以基于点、线要素进行提取的二元化视角，虽具有较高空间定位精度，但无法全面表达微地貌单元的个体特征。本项目拟以高精度DEM，高分辨率遥感影像为数据源，综合数字地形分析、面向地理对象影像分析、深度学习等理论与方法，构建面向黄土微地貌的精细化提取模型。主要研究（1）以地形信息驱动的多尺度分割方法和优化策略，提升面向微地貌单元多尺度分割的精度；（2）基于深度卷积神经网络进行地形信息的多层次深度特征的构建与融合；（3）在此基础上通过多样区的实例验证提取模型的有效性。本项目的开展将拓展 DEM 数字地形分析的方法体系，并可望对黄土高原生态修复与水土保持治理提供技术支撑。

Loess micro-landform represents the spatial distribution and dynamic of matter and energy on the ground surface at a small scale. It is comprehensive since it can reflect the geomorphological evolution and their corresponding interaction. The extraction of it can provide the basic background for the studies on soil erosion, soil and water conservation, geomorphological evolution and so on. However, limited by current methods and technologies, most loess micro-landform extraction methods are based on a binary perspective of extracting point and linear elements, which is hard to reflect the characteristics of loess micro-landform. Taken high-resolution DEM and remote sensing images as data sources, this project intends to integrate digital terrain analysis, geographic object-based image analysis, deep learning and other theories and methods to build a refined extraction model for loess micro-landforms. The project aims to (1) propose a multi-scale segmentation method and optimization strategy driven by terrain information to improve the accuracy for micro-landform object segmentation; (2) construct and fusion of multi-level deep features of terrain information based on deep convolutional neural network; (3) validate of the above extraction model by multiple test areas. This project will extend the methodology of digital terrain analysis and provide technical support for ecological restoration and soil and water conservation in the Loess Plateau of China.