Research On Intelligent Classification Method Of Remote Sensing Image For Land Cover Types Based On Google Earth Engine Cloud Platform

Took global land cover products MCD12Q1 Version 6,Landsat-8 OLI,and Sentinel-2 L1 C as the main remote sensing images,mixed and multiple land cover types with weak independence in the study area were extracted automatically and accurately by combined the SRTM V4.1 auxiliary dataset and thematic indexes to construct a multidimensional classification feature set of the spectrum,texture,and terrain,which was a large training sample dataset derived from Landsat-8 OLI and Sentinel-2 L1 C remote sensing images were automatically created from the global land cover product MCD12Q1 Version 6 IGBP?LC?Type1.And the potential misclassified pixels of land cover types of Landsat-8 OLI and Sentinel-2 L1 C images were reduced through spatial filtering,homogeneous pixel screening,and quality control measures to improve the accuracy of automatic creation of large training samples datasets.Furthermore,the deep residual network ResNet-101 classifier in the convolutional neural network was used to learn the classification rules of the training data of the given category to classify the unknown data and compared with the most commonly used Random Forest(RF),Classification and Regression Tree(CART),and Support Vector Machine(SVM)in traditional classifiers after parameter selection and optimization,method on the reliability of automatic collection of large training sample dataset and the ResNet-101 machine learning algorithm through quantitative evaluation were verified.In the meantime,an intelligent classification method of remote sensing images with fast,efficient,and precise was proposed to weaken human interference and greatly simplify the traditional classification process.This study aims to enrich and improve the extraction theory,classification algorithm,classification accuracy,and analysis and processing platform of Land Use and Cover Change information in remote sensing images,and provide a reference for the effective use and real-time monitoring of land resources.The main steps of the method proposed in this study were all implemented in the Google Earth Engine(GEE)cloud platform to quickly extract and analyze the land cover information of remote sensing images,promote the automation of operating procedures by self-programming independently,integrate data acquisition and processing,and solve the time-consuming and labor-intensive difficulties in traditional land cover classification.The following are the main works and findings:(1)Proposed an image classification optimization system based on parameter adjustment of traditional muti-classifiersAccording to the characteristics of remote sensing images,an image classification optimization system of RF,CART,and SVM by kernel parameter adjustment was proposed.Moreover,the applicability and superiority of the optimized parameter were proved,and the local search limitation of the classification index function could be broken away.Both Landsat-8 OLI and Sentinel-2 L1 C images showed that the selected and optimized classifier parameters proposed in this study could further refine classification results and improve the accuracy.(2)Constructed a deep-layer ResNet-101 model based on the TensorFlowBy introducing residual attention mechanisms,the goal of eliminating network redundancy and enhancing salient features was achieved.To avoid the disappearance of the gradient caused by the improvement of network layers in the ResNet-101 model,a shortcut was added to the feed-forward neural network to skip one or more layers and merge with the main path at an asynchronous length,and by adding linear projection to ensure that the input and output dimensions were the same.Through hyperparameter settings,the training time and convergence accuracy of the ResNet-101 model in the training dataset reached the optimal value.Finally,the calculation process of the ResNet-101 model based on TensorFlow was expressed in a flowchart,and a specific implementation method for the complete construction and optimization was proposed.Among them,the Re LU activation function was used to improve the convergence speed,and the convolution module with 1×1 was added to change the dimension,thereby improving the classification accuracy.(3)Accomplished the automatic extraction and creation of a large training sample dataset of land cover types based on the global MCD12Q1 land cover productTook global land cover products MCD12Q1 Version 6,Landsat-8 OLI,and Sentinel-2 L1 C as the main remote sensing images,mixed and multiple land cover types with weak independence in the study area were extracted automatically and accurately by combined the SRTM V4.1 auxiliary dataset and thematic indexes to construct a multidimensional classification feature set of the spectrum,texture,and terrain,which was a large training sample dataset derived from Landsat-8 OLI and Sentinel-2 L1 C images were automatically created from the global land cover product MCD12Q1 Version 6 IGBP?LC?Type1.And the misclassified pixels of Landsat-8OLI and Sentinel-2 L1 C images were reduced through spatial filtering,homogeneous pixel screening,and quality control measures to improve the accuracy of automatic creation of large training samples datasets.Finally,an intelligent classification method of remote sensing images with fast,efficient,and precise was proposed to weaken human interference and simplify the traditional classification process.(4)Verified the reliability of the intelligent classification method of land cover through quantitative evaluationThe accuracies of the classification results under the RF,CART,SVM,and ResNet-101 classifiers were quantitatively evaluated.The quantitative evaluation was analyzed based on the confusion matrix between the classified products and sample points.Based on the generated confusion matrix,the classification accuracy indicators such as producer's accuracy,user's accuracy,overall accuracy,Kappa coefficient,omission error,and commission error were respectively calculated.The evaluation results showed that the ResNet-101 model based on Landsat-8 OLI and Sentinel-2L1 C images had both the best classification accuracy and overall performance.(5)Design of self-programming language based on GEE cloud processingAll images were obtained online from the GEE cloud platform.The imported dataset was superimposed with the global vector map or satellite map to form a highly visual interface of data presentation and interactive analysis.The self-programming language was directly coded through the massive datasets and algorithms stored in the GEE cloud platform.The results proved that the method of integrated processing and parallel computing greatly improved operating efficiency and saved local storage space.The powerful cloud computing capabilities of the GEE provided an extremely convenient measure for processing and analysis of images in a large-scale region.