Research On Land Use Classification Of Buffer Strip Based On Deep Learning

With the implementation of China’s ecological civilization construction and sustainable development strategy,the protection and restoration of river and lake ecosystems have become important national tasks.The buffer stripis is an important ecological space for rivers and lakes,plays a significant role in blocking or slowing down direct human interference,protecting biodiversity,and reducing non-point source pollution entering rivers and lakes.To carry out ecological restoration in the buffer strip and protect it,a detailed investigation of land use and coverage in the riparian zone is necessary.Traditional manual field surveys are time-consuming,labor-intensive,and have low survey granularity.Similarly,using visual interpretation of high-resolution remote sensing images consumes a significant amount of time and manpower,making it difficult to conduct high-frequency continuous surveys.In recent years,deep learning-based image processing techniques have developed rapidly,and existing research has shown that deep neural network models can be applied to high-resolution remote sensing land use classification tasks with good accuracy.Automated land use classification of buffer strip based on deep learning has become possible.This study has constructed a deep learning-based land use classification model for buffer strip,achieving high-precision and high-efficiency classification of land use types in the buffer strip.The main research results include the following aspects:(1)The remote sensing data set suitable for the task of land use classification in buffer strip.Based on the land use classification standard of land survey,this thesis designs the land use classification standard of buffer strip according to the task requirements,and constructs the land use classification data set of buffer strip according to the standard.Through experimental comparison and verification,according to the data set proposed by this standard,in the model trained by deep convolution neural network,the best classification accuracy and classification effect can be obtained,which is more suitable for the task of land use classification in buffer strip.(2)The land use classification method of buffer strip based on migration.The proposed method can effectively improve the accuracy and generalization ability of the model,and reduce the scale of network parameters and computational overhead.Through the experimental comparison,it is proved that the weight file trained by Love DA data set is used as the input of pre-selected training model to use the network trained by buffer land use classification data set for migration learning,and the prediction results of the model are higher than those without transfer learning.The average miou of the model using transfer learning is 1.08% higher,and the size of parameters is reduced by 67%.Among them,HRNet got the best prediction result on the data set,and the miou was 82.22% and the MPA was 83.9%.(3)Improved HRNet deep neural network COHRNet for land Use classification in buffer strip of rivers and lakes.Based on the optimization of the HRNet network model,COHRNet adds a location attention mechanism to the output of low-level semantic branches to strengthen the features of long-distance ground-object relationships and distinct features of land types.At the same time,a self-attention module is added after HRNet to optimize the junction of land objects.Pixel classification accuracy.Comparative experiments show that,the COHRNet based on transfer learning achieves83.36% miou and 84.1% MPA on the land use classification data set of river-lake buffer strip,which is 1.14% and 0.2% higher than HRNet respectively.The AP of each type of figure is higher than that of HRNet.Compared with HRNet patches,the actual classification results are more complete,richer details and clearer edges.