Research On Classification Method Of Coastal Features In High-resolution Remote Sensing Images Based On Deep Learnin

The advantages of using Earth observation remote sensing imagery to obtain highprecision and large-scale surface features make it an important data source for investigating the use of coastal and marine areas.However,due to the complex and rich spectral features of objects in high-resolution imagery,there are phenomena such as "same object with different spectra" and "different objects with same spectra," which make high-precision classification and recognition of coastal objects a challenge.Therefore,this paper proposes two fine-grained coastal object classification methods suitable for high-resolution imagery using deep learning techniques.The main research contents are as follows:(1)This paper proposes a fine-grained coastal object classification method for highresolution imagery based on multi-level features.Firstly,the VGG19 convolutional neural network is used to extract spatial pattern features of remote sensing imagery.Secondly,genetic algorithms are used to optimize random forest parameters for pixel-based and spatial-pattern-based classification models.Then,the D-S evidence theory is used to fuse the classification models to achieve preliminary classification of coastal objects.Finally,based on the preliminary classification results,spatial contextual information of different object patches is extracted,and particle swarm optimization is used to optimize support vector machine parameters for secondary classification of remote sensing imagery objects.Experimental results show that the overall accuracy and Kappa coefficient of the proposed method can reach 97.21% and 0.9642,respectively.The proposed method provides a feasible and effective approach for extracting coastal object classification,while reducing the training time and ensuring high classification accuracy.(2)A method for automatic extraction of ground features from salt fields and aquaculture ponds based on improved UNet has been proposed.Firstly,in order to match the model complexity with the sample complexity,the model reduced the network layers of the traditional UNet semantic segmentation network in its design and added a CBAM module at the skip link.This strategy effectively refined the feature mapping and improved network performance;Secondly,a residual learning module is added at the end of the model encoder to learn residual features;Finally,in order to deal with the overfitting problem caused by the imbalance of data samples,a structured random deactivation module is added after the last two modules of the encoder.To improve the efficiency of model extraction,this article collected Gaofen No.1 and Gaofen No.2 datasets covering the entire coastal zone area of Dongying City in 2017 and 2022,and conducted identification and extraction of two types of land features: salt fields and aquaculture ponds.The extraction results clearly showed the differences in the distribution of seawater aquaculture and salt fields between 2017 and 2022.The experimental results show that this method can outperform various traditional semantic segmentation networks in terms of extraction accuracy and has certain practical value.(3)Designed and developed an automatic coastal feature extraction system suitable for high-resolution images.The system includes functions such as image retrieval,data preparation,model training,and model prediction.Obtaining extraction results from user input image data greatly reduces labor costs and time consumption compared to traditional field survey methods,and can be effectively applied to the actual extraction of coastal features.