Sample Selection Strategy Of Taihang Mountain Land Cover Classification Based On Typical Topographic Factors

The surface cover data produced based on remote sensing image classification technology have important application values in the fields of land resource monitoring and environmental assessment.The topographic effect caused by topographic relief aggravates the phenomenon of "Different spectra for the same feature" and "Same spectrum with different features" in mountain images,which limits the accuracy of mountain cover classification.And the training samples as one of the important factors to determine the supervised classification accuracy,there is a lack of research related to the number and combination of samples in different slope and slope direction regions with large differences in spectral distortion.Therefore,it is of research significance and application value to propose a mountain sample selection strategy for the improvement of mountain cover classification accuracy.This paper takes the sample selection strategy of mountain surface cover classification based on topographic factors typical topographic factors as the research objective,uses Sentinel-2 remote sensing images as the basis,and extracts the slope direction and slope dataset of the study area based on 30 m DEM data while studying the influence of topographic correction on image classification,taking the existing slope and slope direction classification standards as the reference and combining the research content with the actual situation of the study area.Support vector machine(SVM)and random forest(RF)classification algorithms are used to investigate the quantitative relationship between the number and combination of training samples and classification accuracy at different slope and slope down.The influence of topographic factors on image classification and the sample selection strategy of mountain surface cover classification are further explored.Based on this,10 m surface cover classification data of Taihang Mountains are generated.The main conclusions of this thesis are as follows:(1)In the experiments on the effect of slope on the selection of classification samples,both random forest and support vector machine algorithms can achieve overall accuracy of more than 92% at 10-20°,20-30°,30-40°and 40°slope levels with topographic correction into the steady state,which is 1.55%-3.83% better than the accuracy of images with only atmospheric correction.The sample sizes required for the random forest to enter the steady state at four slope levels are 300,1000,1100,and 600;the sample sizes required for the support vector machine are 400,1000,1100,and 800.The sample sizes required for the two classifiers are similar at different slope levels,and both show a trend that the sample sizes required for slopes below 40 increase with slope level,suggesting that the sample sizes required for slopes below 40 increase with slope level.The influence of the topographic effect on the number of training samples is greater than that of the complexity of the ground cover,and the influence of the topographic effect on the number of training samples dominates in this interval.(2)In the experiment on the effect of slope orientation on classification samples,the highest accuracy of random forest in topography-corrected shaded slope and sunny slope images was 88.13% and 87.77%,and the sample size required to enter the smooth state was 900 and600;the highest accuracy of support vector machine was 86.26% and 80.59%,and the sample size required to enter the smooth state was 700 and 400.compared with the accuracy of images with only atmospheric correction.The accuracy was improved by 1.13%-3.20% compared to the image with only atmospheric correction.The sample size required for both classifiers is larger for the negative slope than the positive slope.The improvement of the overall accuracy of the pure shaded slope and sunny slope samples compared with the mixed shaded slope and sunny slope samples is less than 1%,which indicates that the sample combination in the sample selection of different slope directions has little influence on the classification results,and we should focus on increasing the proportion of shaded slope samples in the sample set.(3)A sample selection strategy based on typical terrain factors is proposed.In terms of sample selection quantity,the number of training samples is 30-50 times the feature dimension for 10-20°slope level,100-120 times the feature dimension for both 20-30°slope level and30-40 ° slope level,and 60-80 times the feature dimension for slope level above 40 °;the number of training samples for shaded slope.The number of training samples should be more than 60% of the total number of samples for positive slopes.The overall accuracy of classification enters a high level and reaches a stable state.Within the training sample set,the number of samples of each category should be in accordance with the approximate proportion of surface cover distribution in the study area.In terms of sample selection,in the area with gentle slope,samples reflecting intra-class differences should be selected;in the area with larger slope,samples reflecting inter-class differences should be mainly selected.A rich and sufficient number of training samples can be formed quickly.(4)The above conclusions were synthesized to generate the surface cover classification data of Taihang 10 m Mountain based on 28-view Sentinel-2 images.The final results show that the overall accuracy is 89.19% and the Kappa coefficient is 0.8638,among which woodland,grassland,cropland,impervious surface and water bodies has good classification accuracy,and relatively speaking,bare land shows more underclassification.The same validation sample was used to evaluate the accuracy of the three sets of classification products Glob Land30,GLC?FCS30 and Esri Land Cover in 2020,in which the overall accuracy of Glob Land30 was66.23% with a Kappa coefficient of 0.8659;followed by Esri Land Cover and GLC?FCS30 with an overall The overall accuracy was 63.33% and 62.21%,and the Kappa coefficient was 0.5215 and 0.5243,respectively,and the biggest difference was the distribution range of woodland and grassland,and through the local cross-validation comparison of the four data types,it was found that the training samples with topographic factors were used to achieve the highest classification accuracy in the Taihang Mountains,indicating that the method was effective in classifying large mountainous areas.The method can still achieve good classification results in the classification of large area of mountain cover.