Deep Learning-Based Change Detection Methods Using Remotely Sensed Image

Change detection based on remotely sensed imagery is a key technology for obtaining change information on the earth’s surface,which has been widely used in many fields,including map revision,disaster assessment,urban planning,environmental conservation,as well as monitoring of wetland,agricultural,and forestry.With the rapid development of deep learning techniques,deep learning-based change detection methods have attracted a lot of attention.However,the complexity of land cover,the multi-source of remote sensing image,the diversity of deep learning model structures,and the lack of training samples make it difficult to detect changes in remote sensing images using deep learning techniques.To improve the reliability of the change detection method,this dissertation mainly studies how to apply deep learning techniques to the change detection task of high-resolution remote sensing images.Based on the summary and analysis of the existing deep learning-based change detection methods,a series of novel deep learning-based methods with high reliability for change detection are proposed in consideration of how to realize the processing of multi-sensor and multi-scale remote sensing images,how to reduce the uncertainty of change types in one-class change detection,and how to reduce the need for training samples.The main work of this dissertation is as follows:(1)Existing deep learning-based change detection methods are thoroughly reviewed.It focues on the state-of-the-art methods,applications,and challenges of deep learning for change detection.Specifically,the implementation process of deep learning-based change detection is first introduced.Then,the general frameworks of deep learning-based change detection methods are reviewed and analyzed systematically.Subsequently,the commonly used unsupervised schemes,network structures,the data from different sensors and its application domains for change detection are described.Finally,the major challenges and prospects of deep learning for change detection are discussed and delineated,including heterogeneous big data processing,the unsupervised scheme,and the reliability.This review will be beneficial for researchers in understanding this field.(2)In applications of change detection,remote sensing images usually have different sources and multiple scales,and the number of pixel-level labeled samples may be insufficient.Based on the idea of transfer learning,a high-resolution remote sensing image change detection method based on a deep feature difference convolutional neural network(FDCNN),is proposed for multi-sensor and multi-scale image processing.The method uses a scene classification model to learn the deep features from various remote sensing images and then composes a network with shared weight to generate a multiscale and multi-depth feature difference map for change detection.The network is trained by a proposed change magnitude guided loss function.And at the training stage,only scene-level labeled samples are needed for deep feature learning and a small number of pixel-level labeled samples are used for feature selection.The method is tested on several datasets from different sensors.And the experimental results show that the proposed method achieves better performance compared with other classical methods,which proves that the proposed method has strong robustness and generalization ability to achieve multi-sensor and multi-scale remote sensing image change detection.(3)Considering the application requirements of one-class change detection and the uncertainty of their change types in practical,a novel deep learning-based change detection method is proposed for landslide extraction,namely CDCNN.It studies and implements the landslide extraction process from the aspects of the design of deep learning model structure,training sample generation strategy,image block processing and parallel processing strategies,object-based change detection,and change instance information extraction,aiming to achieve an intelligent,fast,and practical landslide mapping method.The experimental results show that the CDCNN model can more effectively and accurately extract landslide changes from various types of changes in ground objects and the post-processing methods also help to remove non-landslide changes,which effectively alleviates the uncertainty of the change types caused by nonlandslide in the results.Compared with other methods,the proposed method achieves better performance,and aids in making better and more robust predictions in large areas,even if the data volume is massive and the land cover is very complicated.Most importantly,the proposed method has much less human interaction and faster speed,in addition to its high accuracy and stability,which are essentially useful in landslide mapping,slope safety management and landslide disaster mitigation.(4)To alleviate the problems of low-likelihood and lack of pixel-level labeled samples in change detection tasks,a novel neural network integrating the multi-instance learning framework is proposed for change detection,namely CDMI-Net.It can learn deep features of change objects from the input image pairs with scene-level label and provide a score map for change pixels to generated binary results.First,using the multiinstance learning framework to enable CNN to automatically learn the deep features of changes from scene-level samples,without the need for pixel-level samples.By filtering unchanged scenes,the low-likelihood problem in change detection task can be alleviated.Then,a two-stream U-Net with shared weight is designed as a deep feature extractor,which can help remove ground objects that have similar characteristics but do not change over time.And it provides higher-dimensional feature vectors used in multi-instance learning for each pixel.Finally,a multi-instance learning pooling with gated attention mechanism is implemented,which provides a score map of change pixels,and a fast level set evolution algorithm is employed to refine the landslide boundary.The experimental results have shown that the proposed method achieves comparable and even better performance on the testing image pairs than all other methods.This means that the CDMI-Net works well and has the value of practical use for landslide mapping.