基于机器学习的沟谷地貌识别模型对比——以黄土高原典型流域为例

探索沟谷地貌空间分布与环境控制特征之间的联系并构建沟谷地貌准确提取模型,对大尺度范围沟谷提取具有重要意义。基于人工提取黄土高原典型流域沟谷地貌样本,结合不同时期的Landsat8 OLI影像数据和DEM数据,建立随机森林模型确定黄土高原沟谷地貌提取最佳影像时期和最佳组合特征,基于最优模型参数,对比其与支持向量机和人工神经网络沟谷提取模型效果,验证模型泛化能力。结果表明:(1)黄土高原沟谷提取的最佳影像时期为12月,最佳组合特征集为Red、Blue、H、SWIR1、PNT、Coastal、GLCM4和NIR;(2)3种方法提取测试区域的沟谷空间分布一致,从定性和定量角度进行比较,随机森林模型提取效果最好,验证样区平均总体精度为80.48％,相较于支持向量机模型和人工神经网络模型分别提高4.00和8.63个百分比;(3)测试区域中,沟谷地貌面积约占总面积的56.91％,且呈现西北至东南方向逐渐集中的特点。研究表明随机森林模型在黄土高原地区高精度沟谷地貌识别研究中综合表现最佳,可大范围推广。

Comparing the Performance of Machine Learning Models for Identifying Gully Landforms-A Case Study of a Typical Watershed on the Chinese Loess Plateau

Exploring the relationship between spatial distribution and environmental control characters of gully landforms and building accurate extraction model are of great significance for gully landforms extraction in large scale. Based on the artificial extraction of gully landform samples combing with Landsat8 OLI image data with different periods of and DEM data of a typical watershed on the Chinese loess plateau, the random forest model was established to determine the best period for gully landforms extraction and the best combination of gullying features. Then, combined with the optimal model parameters, results of random forest were compared with support vector machine and artificial neural network model to validate the model generalization ability. Our results showed that: (1) The best image period for gully extraction was in December, and the best combination feature set was Red, Blue, elevation (H), SWIR1, positive and negative terrain (PNT), Coastal, texture (GLCM4) and NIR; (2) The distribution of gully landforms in the testing area extracted by three methods had consistently spatial pattern. Based on qualitatively and quantitatively modelling performance, the random forest model presented the best extracting performance, with the average overall accuracy of 80.48%, which was higher by 4.00 percentage and 8.63 percentage compared with the support vector machine model and the artificial neural network model, respectively; (3) The gully landforms accounted for 56.91% of the total testing area and the distribution of gullies in testing area was gradually concentrated from northwest to southeast. The results show that the random forest model has the best comprehensive performance in the study of high-precision gully landforms identification on the Chinese Loess Plateau, and can be widely extended.