中文摘要：

地貌形态特征识别与分类,对生态环境、水文研究及地质构造分析等地学研究具有重要意义,已成为现代地貌学的一个研究热点。传统的统计模式识别方法精度较低,难以解决线性不可分的模式分类问题。人工方法虽然识别精度高,但因各人认知偏差导致的识别误差难以控制。人工神经网络作为一种动态信息处理系统,能有效解决线性不可分的地貌类型识别问题。坡谱是利用微观地形定量指标来反映宏观地形特征的有效方法,在地貌学研究中正受到广泛的关注。本文以陕北黄土高原8个不同地貌类型区的数字高程模型（DEM）为实验数据,以流域为分析单元,提取坡谱及其特征指标作为描述地形特征的定量因子,并通过BP神经网络的构建与学习,进行黄土地貌类型自动识别。实验结果表明,在8种地貌类型的样本数据中,第1次实验正确识别率平均值达70%;第2次和第3次实验中,去除相似度较高的峁状丘陵沟壑或峁梁状丘陵沟壑任一种地貌类型后,正确识别率平均提升为80%和85%。经Kappa系数验证,该方法能以DEM数据有效识别不同类型的黄土地貌。

英文摘要：

As a research hot-spot of modern geomorphology, landform recognition and classification are important in various study areas such as ecological environment, hydrology and geological structure analysis. Traditional recognition methods, which are inadequate to solve the linear inseparable problem of pattern recognition, exhibit a low accuracy in landform recognition. As a dynamic information processing system, neural network is capable to deal with linear inseparable in landform recognition. Slope spectrum is an effective method to reflect the macro terrain features with quantitative micro-terrain-indicators. It has been receiving widespread attentions in geomorphology. This paper introduces an automatic recognition method based on slope spectrum and neural network. Using DEM data of eight sample areas with different loess landform types in Shaanxi Province, ten small watersheds and their slope spectrums are extracted for each of the eight sample areas. Then, we calculate the slope spectrum indices of these eighty small watersheds and use the indices to construct BP neural network for loess landform automatic recognition. Among the eighty small watersheds, 60% of them are randomly selected as training samples and 40% of them are selected as verification samples. Recognition results show that the accuracy rate is 70% on average for the eight sample areas, and it would be raised to 80% or 85% when the landform types of Loess Hilly-gully or Loess Hill-ridge are eliminated from the eight sample areas respectively. This study indicates that slope spectrum is capable of handling the linear inseparable problem in landform recognition.