Deep Learning Based Methods For Remote Sensing Image Classification,Object Detection And Extraction

Over the past few decades satellite remote sensing images have recorded various kinds of information from the earth's surface and have been widely used in studies of climate,agriculture,forestry,hydrology,urban planning,national security and other key areas.It is crucial to automatically extract relevant information from the large number of remote sensing images and improve the accuracy of remote sensing image analysis.In recent years,artificial intelligence,especially deep learning techniques,have become increasingly popular and have had a great influence on the remote sensing image analysis domain.Convolutional neural networks(CNN)based deep learning methods have shown great potential in land cover classification,object detection and extraction tasks in remote sensing domain.However,existing studies still have limitations in many aspects,e.g.,the scale of the study area,the fusion of multiple data sources,the accuracy of the method,etc.This study aims at interpreting the structured information of the remote sensing images intelligently based on deep learning.More specifically,we focus on solving three typical geoscience research issues,i.e.,land cover classification,oil palm tree detection,and building extraction.We make substantial improvements to existing studies in terms of dataset construction,method design,results analysis,etc.First,we propose a deep learning-based approach for land cover classification and mapping in China.As the first deep learning-based method evaluated by samples cov-ering the whole of China,we combine high-resolution Google Earth imagery with 30m-resolution data(e.g.,Landsat data).We design and propose a novel deep CNN based land cover mapping approach which takes full advantage of different data sources in differ-ent spatial and spectral resolutions.Experiment results show that our proposed method achieves a classification accuracy of 84.4%on the whole validation dataset in China,im-proving the previous state-of-the-art accuracy by a further 4%.Moreover,our proposed method significantly reduces misclassifications between different vegetation types and the impervious type.Second,we propose a deep learning-based oil palm tree detection approach,which is the first attempt to apply deep learning methods to satellite image-based tree crown detection.We build an oil palm detection dataset using a Quickbird satellite image.We design a two-stage convolutional neural network(TS-CNN)based approach and an end-to-end approach for oil palm tree detection.Experiment results show that our proposed method achieves an overall F1-score of 94.53%in a 55 km~2study area in Malaysia,improving the F1-scores obtained from existing methods by 7%-16%.Moreover,there is little confusion between oil palm trees and other vegetation in the whole image results obtained from our proposed method.Third,we propose a semantic segmentation-based building extraction method.To the best of our knowledge,this is the first effort to explore the combination of multisource GIS map datasets and multispectral satellite images for building footprint extraction.We propose a complete solution for building extraction based on the U-Net semantic segmentation model.The proposed method obtains the total F1-score of 70.4%on the validation dataset and improves the total F1-score obtained from the winning solution of the 2017 Space Net Building Detection Challenge by another 1.1%.The final building extraction results are comprehensively analyzed based on the actual situation of four cities.Our proposed methods effectively improve the results of large-scale land cover map-ping,tree detection and building extraction,contributing to more accurate fundamental data for ecology,agriculture,forestry,urban planning,and other geoscience research domains.