基于BP神经网络的黄土侵蚀沟自动提取方法

Automatic extraction method of loess erosion gullies based on BP neural network

黄土侵蚀沟信息是研究沟壑地貌土壤侵蚀的重要依据,而目前遥感影像提取方法中存在沟沿陡边的遮挡问题,由此,本文研究了一种基于反向传播神经网络的自动提取方法。首先,利用数字高程模型,基于黄土侵蚀沟的特征分析,选取横向坡度、坡度变率、坡向变率、地形起伏度、地表切割深度、高程变异系数和地表粗糙度作为地形特征因子,通过沟谷网络的计算,制作黄土侵蚀地貌的训练样本数据集;然后,基于反向传播神经网络模型的训练实验,选择Trainbr作为神经网络模型的学习算法;最后,应用侵蚀沟地貌的神经网络模型对测试数据集进行提取实验,并与随机森林和支持向量机方法进行比较。结果表明,本文方法的准确率好于其他方法,漏分情况相对较少,可以满足黄土冲沟信息的高效与准确提取需求。

Whether it pertains to soil and water conservation and management across the loess plateau or the formulation of scientific plans for ecological preservation within the Yellow River Basin,clarifying the spatial distribution and changes associated with loess erosion gullies remains imperative.Despite the existence of prior research regarding the extraction of loess erosion landforms,the accuracy and efficiency of erosion gully extraction processes continue to be unsatisfactory.This paper takes the loess erosion landform in Ansai District,Yan'an China,situated in the upper reaches of the Yellow River,as a case study.The study focuses on a method of extracting loess erosion gullies based on BP neural network,with the goal of achieving automated extraction of loess erosion gully landform.The method uses digital elevation model(DEM)data as its foundational dataset and selects lateral slope,slope variation rate,aspect variation rate,terrain undulation,surface cutting depth,elevation variation coefficient,and surface roughness as terrain features.These features are selected through an analysis of the characteristics of loess erosion gullies.To produce the training sample data of loess erosion landform,the gully network is extracted based on the Hydrology principles.Subsequently,the thalweg line is extracted using the slope distortion neighborhood judgment method.By considering the spatial relationship between the gully line and the thalweg line,pseudo gullies are removed to obtain a relatively accurate gully network,which serves as the basis for selecting gully samples.This study selects three sample watersheds within the research area,utilizing the pixels in two of them as training samples.The two training samples have a total of 21897 points,including 11897 gully points and 10000 non gully points.Using one watershed as the test sample,a total of 12087 points were identified,including 8087 gully points and 4000 non gully points.The network layer of the BP neural network structure is designed as 3 layers.Through training experiments of the BP neural network model,considering F1 score,accuracy,and time-consuming performance,the function Trainbr is selected as the learning algorithm of the BP neural network model.After evaluating F1 scores and accuracy corresponding to different number of neurons,set the number of neurons in the hidden layer to 10,and select the hyperbolic tangent sigmoid function TanSig as the transfer function for both the hidden layer and the output layer.The neural network model of erosion gully landform is applied to extract loess erosion gully landform from the test data set,and its performance is compared with Random Forest and Support Vector Machine methods.Through the comparison of test sample extraction results,it is evident that all three models have missed points when compared to actual gully network data.However,the automatic extraction model based on the BP neural network performs better in the erosion gully test set of the research area than other methods,with better continuity and integrity of recognition results and relatively fewer missed points.The overall accuracy of the erosion gully automatic extraction model,utilizing the BP neural network,attains 93.60%,accompanied by an F1 score of 0.95.The method proposed in this paper yields the most effective results for extracting erosion gullies within the study area.The results indicate that the erosion gully feature factor proposed in this paper can better reflect the terrain characteristics of erosion gullies,and the constructed BP neural network model can effectively achieve automatic extraction of erosion gullies in the research area.