Deep Learning Based Terrace Information Extraction And Its Impact On Soil Erosion Estimation

In order to control soil erosion on the Loess Plateau,the state has implemented a series of soil and water conservation measures,including the conversion of slopes into terraces.As an important soil and water conservation measure in the Loess Plateau region,terracing is of great significance to the prevention and control of soil erosion because of its efficient water,soil and fertilizer retention.Traditional terracing statistics methods are slow and low in accuracy,and there is an urgent need to propose a fast and effective terracing extraction method to scientifically and accurately assess the impact of terracing on soil erosion estimation,and provide a data basis for soil and water conservation work and ecological environment construction on the Loess Plateau.This paper takes the Yanhe River basin as the study area,based on the high-score one data(GF-1),uses the deep learning method and object-oriented classification method to extract terraces information respectively,integrates GIS;RS,RUSLE model and other techniques to quantitatively evaluate the terraces’ effects.The main research results are as follows.(1)Based on the deep convolutional neural network model,a combination of the backbone feature network and three loss functions is explored for optimal extraction of terraces.The results show that the terraces are optimally extracted using the ResNet50 backbone network and the Lovasz-Softmax loss function under the PSPNet model.the average cross-merge ratio of the ResNet50 backbone network is 6.19%,5.51%and 13.35%higher than that of the Mobilenet backbone network in the training,validation and test sets,respectively,and the accuracy is improved by 5.67%,5.36%and 12.37%respectively.Therefore,the extraction accuracy of the ResNet50 backbone network is significantly better than that of the Mobilenet backbone network.Based on the combination of ResNet50 backbone network with three loss functions,Log-softmax,CrossEntropyLoss and Lovasz-Softmax,respectively,the terrace extraction accuracy was analysed,and the results showed that using the Lovasz-Softmax loss function was significantly better than Log-softmax and CrossEntropyLoss loss functions by 6.23%and 3.41%,and the accuracy by 2.73%and 2.12%,respectively.The experimental results show that the ResNet50 backbone network using the PSPNet model and the Lovasz-Softmax loss function is the best strategy for extracting terrace information.(2)The accuracy of the deep learning method and the object-oriented classification method in extracting terrace boundaries were compared,and the results showed that the deep learning method was more accurate in extracting terrace boundaries.Based on the spectral,textural and morphological features of the terraces,the classification rules of the terraces were constructed,and the positive judgment rate and average intersection ratio of the extracted terraces were 81.25%and 78.31%respectively;the positive judgment rate and average intersection ratio of the PSPNet-based model in terrace information extraction were improved by 9.00%and 9.01%respectively compared with the object-oriented classification method.Therefore,the extraction of terrace information using deep learning methods is more accurate.(3)The terraced fields in the Yanhe River Basin were extracted using a combination of deep learning optimal strategies to quantify the effect of terraces on soil erosion estimation under different classification results.terraces in the Yanhe River Basin covered an area of about 307.56 km2 in 2015.The contribution of terraces in reducing soil erosion was quantitatively estimated using the RUSLE model,and the results showed that in 2015 the soil erosion in the Yanhe basin without considering terraces was 76.84×106t,and the soil erosion after considering the role of terraces was 69.77×106t,and the terrace information influenced about 9.20%of soil erosion;after considering terraces,the percentage of slightly eroded,lightly eroded and moderately After considering the terraces,the area share of slight erosion,light erosion and moderate erosion increased by 0.08%,2.36%and 1.03%respectively,and the area share of strong erosion,very strong erosion and severe erosion decreased by 0.05%,0.83%and 2.69%respectively,and the construction of terraces could effectively reduce the amount of soil erosion.In addition,the effects of different classification accuracies on soil erosion estimation were calculated.The difference in soil erosion assessed by terrace information extracted based on deep learning and terrace information extracted by object-oriented classification method amounted to 3.29×104t.Therefore,accurate extraction of terrace information has an important impact on the assessment of soil erosion in the Yanhe river basin.