Figure 4 - uploaded by Di Yang
Content may be subject to copyright.
Source publication
In this article, we propose a novel difference image (DI) creation method for unsupervised change detection in multi-temporal multi-spectral remote-sensing images based on deep learning theory. First, we apply deep belief network to learn local and high-level features from the local neighbour of a given pixel in an unsupervised manner. Second, a ba...
Similar publications
In this paper, a high precision detection method of salient objects is presented based on deep convolutional networks with proper combinations of shallow and deep connections. In order to achieve better performance in the extraction of deep semantic features of salient objects, based on a symmetric encoder and decoder architecture, an upgrade of ba...
Using CNN deep learning to learn about fMRI data. Very impressive results.
Citations
... Concerning the methodology, the state-of-the-art in machine learning on satellite data demonstrates that machine learning is the most appropriate and most successful approach to image recognition (LeCun et al., 2015;Zhang et al., 2016;Cao et al., 2017;Cheng et al., 2017;Zhu et al., 2017). In this chapter, rather than illustrating how to develop indicators through digital image processing, we will focus on after-processed remote-sensing indicators as available data products, which can be employed for studying human migration and humanitarian aid. ...
With the rise of Big Data use, deployment of earth observation and remote sensing has gained popularity also for migration studies. Not only the impact of environmental factors, but also of conflict zones and disaster emergence for human mobility and migration can be investigated through analysing different types of remote sensing data by computational methods. In this chapter, after construing the essential concepts of remote sensing, sensors and satellite systems and highlighting the data content and access to publicly available remote sensing data products, we illustrate three application-specific show cases. These are on ethnic migration, environmentally induced human migration and armed conflict related forced migration from Uganda, Mexico, and Colombia, respectively. With this chapter, we documented that satellites are a high potential data source for deriving population estimations for at-risk-communities, displaced people, and refugee groups, for predicting forced migration flows, notably in the context of natural disasters, climate extremes and conflict zones; as well as for generating essential information to facilitate assistance and support for pre- and during-crisis/disaster resilience-building interventions. We conclude that further cooperation between scientists and practitioners is crucial to enhance the benefits and to address the challenges for developing human-rights compliant applications.
... Compared to Multispectral Images (MSIs), HSI has highresolution spectral and spatial information with hundreds of bands, optimizing the quality of HRS images [4]. Since HSIs include a plethora of spectral information, they have been extensively used in various domains, including image classification [5], [6], anomaly detection [7], [8], change detection [9], [10], and so on. ...
In recent decades, the wide use of deep learning-based methods has consistently improved the performance of remote sensing images and is widely used for hyperspectral change detection (HCD) tasks. However, most of the existing HCD method is based on the convolutional neural network (CNN), which shows limitations in long-range dependencies and also cannot mine sequence features well. The CD performance still has margins for improvement. In this study, inspired by the excellent performance of transformers in computer vision and which has shown a significant ability to model global dependencies to attenuate the loss of long-range information, we built a hybrid spatial-spectral convolutional vision transformer (SSViT) for HCD. Our proposed method combines the merits of CNN and transformer to fulfill effective and efficient HCD. This study focused on highly reliable pseudo-sample data generation by selection scenario. To generate a pseudo sample, we have used different methods: (1) we predict change and no-change areas by using Euclidean distance, (2) thresholding by Chan-Vese segmentation method for determining change and no-change pixels for intensity maps, (3) sorting of change and no-change pixels, and (4) selection of the minimum value of initial no-change pixels as pseudo change sample data, in addition to, choosing the maximum intensity value for change candidate pixels as change sample data. The highly reliable change pixels were selected, and then pseudo-training data was used to train the SSViT model. At last, the change map is generated by training the SSViT network based on pseudo-training data. The performance of the SSViT model is evaluated for real-world hyperspectral (HS) datasets with different change landcover types. Furthermore, a new series of HS images is introduced for CD purposes. The results of CD show that the HS images have a high potential for detecting subtle changes. The experimental results demonstrate that the proposed SSViT could outperform the advanced HCD methods.
... For acquiring the ability to solve complex problems, the restricted Boltzmann machine (RBM), as a two-layer structure with the visible layer and the hidden layer, reduces the dimension of complex RS data through the neuron association between layers. Deep belief networks (DBN) [17] are stacked by multiply RBMs, its independent hidden layer units are training separately with the joint probability distribution of the data. Experiments show that DBN automatically acquires the abstract change information hidden in the data, which is difficult to interpret, and improves the assimilation effect of the unchanging areas, meanwhile highlighting the change. ...
... Several deep-learning methods for CD have been proposed. In [8], a method was proposed to compute the difference image using a backpropagation algorithm and a deep belief network. ...
This study aimed to adapt Convolutional Neural Networks (CNN) to solve the problem of change detection using remote sensing imagery. Specifically, the goal was to investigate the impact of each CNN layer to detect changes between two satellite images acquired on two different dates. As low-level CNN layers detect fine details (small changes) and higher-level layers detect coarse details (large changes), the idea was to assign a weight to each layer and use a genetic algorithm based on a training dataset to generalize the detection process on the test dataset. The results showed the effectiveness of the proposed approach based on two real-life datasets.
... Algunos trabajos también incluyen el uso de metodologías hibridas. Los principales procesos de detección de cambios implementando DL se basan en diferencia de imágenes (DI) [25]. Una de las grandes ventajas del uso de redes neuronales profundas es la capacidad de emplear y analizar datos de múltiples fuentes (Ópticas, Radar, DEM, etc.) y de diferentes periodos de tiempos combinando varias técnicas de fusión, lo cual ayuda a mejorar la exactitud de los procesos de detección de cambios [26]. ...
The processes of change detection (CD) in land use and land cover through satellite images acquired in different temporal phases represent a key process to monitor the land, the environment and evaluate disasters. Currently the methodologies used for CD based on pixels do not allow processing a large amount of images massively, so it is not possible to analyze all the available information. Main aim of this research is to evaluate the performance of Deep learning methods for CD, grouped in two approaches, across open mining areas from PlanetScope (PS) image time series of the Cerrejon mine in Colombia. Two approaches for change generation are proposed, one based on post-classification comparison in which two convolutional neural network (CNN) architectures are evaluated: U-Net and Feature Pyramidal Network (FPN) for the classification of mining areas along the time series, for this purpose different models with different hyperparameters were generated and trained to select the most suitable for such process and subsequently perform the difference between the periods of the series; and a second approach based on direct change detection in which a modified U-Net network was evaluated from pairs of images. For this purpose, different models were also trained selecting the most appropriate for the detection of changes for each period of the time series, obtaining a map of changes for each approach, each one the results were validated; the most appropriate approach was number 2 (Direct change detection), with kappa accuracies greater than 0.9 in each period of the time series.
... For example, Chen, Shi, and Gong (2016) used the difference between bi-temporal deep features and an initial DI to design a cost function. Cao et al. (2017) extended a backpropagation algorithm to build a cost function between two DBNs. ...
A large number of publications have incorporated deep learning in the process of remote sensing change detection. In these Deep Learning Change Detection (DLCD) publications, deep learning methods have demonstrated their superiority over conventional change detection methods. However, the theoretical underpinnings of why deep learning improves the performance of change detection remain unresolved. As of today, few in-depth reviews have investigated the mechanisms of DLCD. Without such a review, five critical questions remain unclear. Does DLCD provide improved information representation for change detection? If so, how? How to select an appropriate DLCD method and why? How much does each type of change benefits from DLCD in terms of its performance? What are the major limitations of existing DLCD methods and what are the prospects for DLCD? To address these five questions, we reviewed according to the following strategies. We grouped the DLCD information assemblages into the four unique dimensions of remote sensing: spectral, spatial, temporal, and multi-sensor. For the extraction of information in each dimension, the difference between DLCD and conventional change detection methods was compared. We proposed a taxonomy of existing DLCD methods by dividing them into two distinctive pools: separate and coupled models. Their advantages, limitations, applicability, and performance were thoroughly investigated and explicitly presented. We examined the variations in performance between DLCD and conventional change detection. We depicted two limitations of DLCD, i.e. training sample and hardware and software dilemmas. Based on these analyses, we identified directions for future developments. As a result of our review, we found that DLCD’s advantages over conventional change detection can be attributed to three factors: improved information representation; improved change detection methods; and performance enhancements. DLCD has to surpass the limitations with regard to training samples and computing infrastructure. We envision this review can boost developments of deep learning in change detection applications. © 2022 Wuhan University. Published by Informa UK Limited, trading as Taylor & Francis Group.
... Cao and others applied the fuzzy classification method to the study of impervious surface performance and compared it with the LSMA method. e results show that the fuzzy classification method is better than the LSMA method in any season, and the performance of the fuzzy classification method is better in both highdensity residential areas and low-density residential areas [6]. Gao and others studied the land use types of Xinqiao town based on quick bird images. ...
In order to study the remote sensing recognition and classification of forest vegetation based on image feature depth learning, this paper presents a deep learning method using classical algorithms such as the maximum class method and maximum entropy method, as well as the FRFCM algorithm and convolutional neural network. In this method, SVM is used to train, classify, and recognize the color information of a high-resolution remote sensing image, remove the nongreen background of the classified image, and finally convert it to HSI space for morphological opening and closing reconstruction, so as to obtain the final extraction target. Then, a visual interface is designed to facilitate operation, which can compare the forest vegetation extraction results and operation processing time under different algorithms, so as to realize the rapid and accurate monitoring of karst forest vegetation change with remote sensing big data. The algorithm research shows that the overall accuracy of multifeature ant colony intelligent classification based on vegetation zoning is 88.85%, Kappa = 0.86, which is better than the traditional remote sensing image classification method, and provides an effective method for land use land cover remote sensing information extraction in large-scale complex terrain areas. In this way, the error extraction and missing extraction can be reduced in the extraction results of forest vegetation area in remote sensing images, and the experimental extraction results will be further close to the optimal segmentation effect.
... Cao et al. [136] developed an approach for generating unique difference images (DIs) for unsupervised CD in multispectral RS datasets. First, they used a DBN to learn local and high-level features from the local neighbour of a given pixel in an unsupervised way. ...
Images gathered from different satellites are vastly available these days due to the fast development of remote sensing (RS) technology. These images significantly enhance the data sources of change detection (CD). CD is a technique of recognizing the dissimilarities in the images acquired at distinct intervals and are used for numerous applications, such as urban area development, disaster management, land cover object identification, etc. In recent years, deep learning (DL) techniques have been used tremendously in change detection processes, where it has achieved great success because of their practical applications. Some researchers have even claimed that DL approaches outperform traditional approaches and enhance change detection accuracy. Therefore, this review focuses on deep learning techniques, such as supervised, unsupervised, and semi-supervised for different change detection datasets, such as SAR, multispectral, hyperspectral, VHR, and heterogeneous images, and their advantages and disadvantages will be highlighted. In the end, some significant challenges are discussed to understand the context of improvements in change detection datasets and deep learning models. Overall, this review will be beneficial for the future development of CD methods.
... Cao et al. [136] developed an approach for generating unique difference images (DIs) for unsupervised CD in multispectral RS datasets. First, they used a DBN to learn local and high-level features from the local neighbour of a given pixel in an unsupervised way. ...
... However, the units within the same layer are not connected and each hidden layer serves as a transparent layer for the next layer. It can be greedily trained, i.e., one layer at a time, and appears in many unsupervised image processing techniques (Cao et al., 2017;Chu et al., 2016;Samadi et al., 2019). A DBN-assisted framework is proposed for the classification of numerous changes from an image (Su et al., 2017). ...
Satellite images taken on the earth's surface are analyzed to identify the spatial and temporal changes that have occurred naturally or manmade. Real-time prediction of change provides an understanding related to the land cover, environmental changes, habitat fragmentation, coastal alteration, urban sprawl, etc. In the current study, various digital change detection approaches and their constituent methods are presented. It was found that (i) change vector analysis method provides better accuracy among the algebra-based change detection approach, (ii) discrete wavelet transformation is better among transformation techniques, (iii) considering the artificial neural network and fuzzy-based approaches to analyze the prediction performance over the traditional state-of-the-art approaches, (iv) analyzing the promising outcomes generated by deep learning techniques for difference analysis related to the images captured at a different instance of time. The brief outlines of different change detection approaches are discussed in this study and addressed the need for improvement in the methods that are developed for the detection of a change in the remote sensing community.
