Research On Rock Information Extraction From Remote Sensing Images Based On Deep Learning

With the continuous promotion of the "One Belt,One Road",there are many transportation infrastructure construction projects such as roads and railways,most of which are far away from China and involve a lot of transportation of road and railway construction materials.If local materials can be used,a lot of capital and labor costs can be saved.At present,in the process of road and railway construction in China,the way of manual survey is usually used to investigate the quarrying area,which has low efficiency and limited scope.Especially in the process of road and railway construction abroad,it is difficult to carry out manual investigation due to the complex environment and lack of necessary basic data involving many countries.Therefore,a set of macroscopic and low-cost survey means is urgently needed to determine the sampling area and the identification and judgment of material characteristics along the construction route.To solve the above problems,this paper carries out experiments and innovations from two aspects: machine learning and deep learning.In object-oriented machine learning classification experiments,the dimension of feature space is often high,leading to "dimension disaster".In this paper,a feature optimization method is proposed to improve the classification efficiency and accuracy of remote sensing image classifier.Specifically,this method combines two algorithms,Fisher Score and mRMR.Firstly,Fisher Score method is used to calculate the importance of the feature indicators,and then mRMR algorithm is used to select the features that are most relevant to the category and least redundant with each other.Through this feature optimization method,the optimal feature subset can be obtained,and the feature subset can be used for automatic classification of remote sensing images.This method can effectively improve the accuracy and efficiency of remote sensing image classification and is suitable for different types of remote sensing image classification such as Gaofen-2 remote sensing image.Experiments show that compared with the traditional mRMR,Fisher Score and Relief F methods,the Fisher Score-MRMR(Fm)method proposed in this paper has higher accuracy of remote sensing image classification.To solve this problem of the semantic segmentation of high-distinguishability remote sensing images,such as discontinuous edge segmentation of ground objects and low accuracy of small target segmentation,this thesis comes up with an optimized remote sensing image segmentation way,which combines ResNet50 and DeepLabV3+ deep learning networks.Compared with the previous trunk network Xception,ResNet50 can better process high-resolution images,and ELU activation function,Adam gradient descent method and Focal Loss loss function can improve segmentation accuracy and convergence speed.In addition,the lightweight MobileNetV2 network structure is used to obtain multi-scale road information features,thus reducing the loss of rock remote sensing image details and improving the rock extraction accuracy of the network model.The experimental results show that compared with the previous ways such as FCN,UNet and DeepLabV3+,the arithmetric can obviously improve the image detection precision and the overall accuracy index,up to 90%.