Large-scale Land-cover Classification With High-resolution Remote Sensing Images

Land-cover classification with remote sensing images plays an important role in many applications such as land resource management,urban planning,precision agriculture,and environmental protection.At present,because of the large ground coverage,low-/medium-spatial resolution remote sensing images are generally used in the practical land-cover applications.However,due to the lack of spatial information,these images are insufficient for detailed mapping in high heterogeneous areas.In recent years,high-resolution remote sensing images are increasingly available.Compared with low-/medium-spatial resolution images,they provide rich texture,shape,and spatial distribution information of ground objects,which contribute significantly to distinguish fine land-cover categories.Nevertheless,because of the narrow ground coverage and high economic costs,high-resolution remote sensing images can usually only be used in classification for some specific small regions.Therefore,to meet the application needs of land-cover mapping,it is highly demanded to study how to exploit high-resolution remote sensing images for large-scale land-cover classification.Image features are essential for remote sensing image classification.Since the information in the high-resolution remote sensing images is highly detailed,the hand-craft features cannot accurately describe the attribute of ground objects.Recently,it has been reported that high-level semantic features can largely improve the classification performance.Among then,deep learning-based methods has drawn wide attention in the interpretation of high-resolution remote sensing images.Although these methods have made some achievements,there are the following problems in applying deep learning to large-scale land-cover classification:(1)Most of existing land-cover datasets are small in size and lack of diversity,in consequence,they cannot fully represent the distributions of real ground objects.It is difficult to design and train deep learning models with strong generalization based on them.(2)Although deep models can learn high-level semantic features from the data,their fixed receptive field limits the capture of information.In high-resolution remote sensing images,the scale of the ground objects varies greatly.For the decrease of separability caused by the scale change,the discriminative ability of deep model is insufficient.(3)Large-scale land-cover mapping usually needs to stitch together the images captured by different sensors from different geo-locations.However,diverse imaging conditions leads to diverse feature distributions,making classification models invalid for different images.Although deep models possess a certain transferability,new specific annotated samples are often necessary for model fine-tuning,which is inefficient when solving practical problemsIn response to the above problems,several issues about large-scale land-cover classification are investigated,including dataset construction,semantic feature representation,and deep model transfer.The main contributions of this thesis are summarized as follows:(1)This thesis constructs a large-scale land-cover classification dataset with high-resolution remote sensing images.The dataset is named as Gaofen Image dataset(GID),which consist of 150 pixel-level annotated Gaofen-2 images.Images of GID are acquired from more than 60 regions in China and cover areas more than 50,000 km~2.GID is the first and largest well-annotated land-cover classification dataset with spatial resolution up to 4 meters.It can provide the research community a high-quality dataset to advance relevant research.(2)This thesis proposes a classification algorithm based on aggregated multi-scale deep features.In order to recognize ground objects with large scale differences,multi-scale deep features are extracted and aggregated for obtaining contextual information.Then,a patch-wise classification is conducted on the image relying on the multi-scale contextual information.Meanwhile,a hierarchical segmentation is used to obtain the object boundary information,which is integrated into the patch-wise classification map for accurate results.(3)This thesis proposed a semi-supervised transfer learning method based on deep semantic association.In order to efficiently classify images captured from different imaging conditions,reliable samples are mined from unlabeled images for deep model fine-tuning.Concretely,a pre-trained deep model is used to assign pseudo-labels to patches of the target image,and the pseudo-labels are used to judge the semantic relevance of patches and existing labels.The patches determined to be reliable are used as re-training samples to realize automatic model transfer.(4)This thesis validates the proposed classification and transfer learning methods on several different remote sensing data,including Gaofen-1,Gaofen-2,Jilin-1,Ziyuan-3,and Sentinel-2A images.The experiments achieve promising results and show the effectiveness of the proposed schemes.Furthermore,land-cover mapping in Wuhan city is implemented using Google Earth platform data,which demonstrates the potential of the proposed methods in practical applications.