Study On Land Cover Classification In Arid Areas Based On Deep Learning Semantic Segmentation Of High-Resolution Images

The northwest arid zone in China is characterized by a complex geographical environment and scarce water resources.The land cover type is predominantly desert,interspersed with oasis patches containing a rich variety of land cover types.Some human activities have destroyed the ecological environment and induced an imbalance in the ecosystem,which affects the survival and development of human beings.A comprehensive grasp of land cover distribution and its changes in the arid zone is of great importance to regional sustainable development,ecological protection,agricultural development,land management,urbanization,and other aspects.Remote sensing technology can provide a large amount of data support for dynamic monitoring in the land cover classification of the northwest arid zone.The traditional remote sensing land cover classification method has problems such as the pepper and salt phenomenon and object segmentation error.In the arid zone,there are phenomena such as uneven distribution of feature types,many similar and easily confused types of features,and smaller target features.The above-mentioned reasons have led to the insufficient results of arid zone land cover classification.It is an urgent problem to improve the extraction accuracy of small target features,especially linear small targets,and the recognition ability of confused features,to improve the overall classification accuracy.The rapid development of deep learning semantic segmentation methods,with advantages such as automatic learning of data features,provides new possibilities for improving land cover classification accuracy in arid regions.Based on the aerial high-resolution image dataset of Minqin County,Gansu Province,China,in this thesis,we propose an improved semantic segmentation model,namely u Seg Net,based on Seg Net and U-Net model,for high-resolution land cover classification in the northwest arid zone.The main results and conclusions of this thesis are as follows:(1)Comparison of extraction effects of six typical deep learning models on highresolution datasets in arid zones.The experimental results of Seg Net,U-Net,Res Net50＿FCN,Res Net101＿FCN,PSPNet,and Deep Lab v3+ models on the Minqin dataset show that Deep Lab v3+ and PSPNet have poor extraction effects on individual features and overall.U-Net,Seg Net,Res Net50＿FCN,and Res Net101＿FCN models are better in some categories.The two models with better overall results are Seg Net and UNet,with mean iou ratio of 78.3% and 74.4% and F1 scores of 87.3% and 84.5%,respectively.Seg Net is more effective in extracting details such as boundary and line intersection,and U-Net is more effective in recovering the connectivity of small linear target features.(2)Construction of u Seg Net,a semantic segmentation model for land cover classification in arid areas.Based on the Seg Net model and the U-Net model,the improved u Seg Net model is obtained by introducing additive connectivity operations to enhance the recognition ability of small target features and the extraction ability of boundary features.In this thesis,we compare the u Seg Net model with Seg Net,U-Net,Res Net50＿FCN,Res Net101＿FCN,PSPNet,and Deep Lab v3+,and the results show that compared with other network models,the u Seg Net achieves better results in overall accuracy,precision rate,recall rate,mean Io U ratio,and F1 score,improved by0.3～2.9%,1.0～14.1%,0.8～16.7%,1.3～11.9%,and 0.9～16.0%,respectively.Compared with the dimensional connectivity operation,all assessment metrics achieve better results,with 0.5% improvement in overall accuracy and precision,3.1% improvement in recall,3.0% improvement in mean Io U ratio,and 2.1% improvement in F1 score.The above experimental results demonstrate that the additive join operation and the maximum index pooling operation of the u Seg Net work together to make the model have better reduction ability,which can enhance the effective use of contextual information,reduce the degree of confusion between categories,and improve the overall accuracy.(3)To evaluate the classification ability of u Seg Net for all kinds of features in arid areas.Among all land categories,except garden land,roads,and artificial excavated land,the classification accuracy of most of the ground features is greater than 70%.The three best-classified land objects are forest land,water area,and cultivated land,and the intersection and union ratios are 93.9%,93.0%,and 90.7% respectively;for cultivated land,forest land,and water,the classification performance is the strongest due to their wide distribution or special spectral information and texture features;for four categories of land cover categories,namely,housing construction area,desert and bare ground,grassland,and structures,the extraction effect of the u Seg Net model is inferior to that of cultivated land,forest land,and water due to the similarity of their spectral texture features with other categories of land cover or the small targets.The recognition effect of the u Seg Net model on the garden land,road,and artificial excavated land is poorer among the ten types of categories,mainly because the spectral texture information is similar to that of the other types of land cover due to seasonal and climatic differences,and the feature information is easily lost due to the small target.However,the Io U ratios are higher than those of other models by 1.4～18.5% for garden land,2.7～35.8% for roads,and 1.5～21.3% for artificial excavated land.It is proved that the additive connection operation in u Seg Net can effectively identify and restore small targets,especially narrow linear land classes and blocky small target features.(4)To evaluate the application capability of the model proposed in this thesis.The effects of the u Seg Net model,traditional pixel-based classification methods,and traditional object-based classification methods are compared on the new phase of remote sensing images acquired from July-October 2017,with 1m resolution.On the one hand,although the growth conditions and spectral information of vegetation and texture features differ under different seasons,which make the application data differ significantly from the training data attributes and thus lead to the application results being less than expected,the Io U ratio of cropland is much higher than that of other land categories,which proves that the confidence level of cropland in the application results is higher.On the other hand,compared with the traditional method using 2017 data for supervised classification,the deep learning model is better for the degree of distinction of confused features and small target extraction.The total accuracy and Kappa coefficient of the application results still perform the highest,with the total accuracy being 9.77% and 12.15% higher and the Kappa coefficient being 0.02 and0.18 higher compared to the two traditional classification methods,respectively.The overall reliability of the traditional land cover classification methods is poor.The results based on traditional classification methods are less reliable.In summary,the deep learning semantic segmentation method has obvious advantages in the application of land cover classification.