In the last decade, the rapid development of deep learning (DL) has made it possible to perform automatic, accurate, and robust change detection (CD) on large volumes of remote sensing images (RSIs). However, despite advances in CD methods, their practical application in real-world contexts remains limited because of diverse input data and the applicational context. For example, the collected RSIs can be time-series observations, and more informative results are required to indicate the time of change or the specific change category. Moreover, training a deep neural network (DNN) requires a massive amount of training samples, and in many cases these samples are difficult to collect. To address these challenges, various specific CD methods have been developed considering different application scenarios and training resources. Additionally, recent advancements in image generation, self-supervision, and visual foundation models (VFMs) have opened up new approaches to address the ”data-hungry” issue of DL-based CD. The development of these methods in broader application scenarios requires further investigation and discussion. Therefore, this article summarizes the literature methods for different CD tasks and the available strategies and techniques to train and deploy DL-based CD methods in sample-limited scenarios. We expect this survey to be able to provide new insights and inspiration for researchers in this field to develop more effective CD methods that can be applied in a wider range of contexts.

Change detection (CD) in high-resolution remote sensing has received large attention due to its wide range of applications. Many methods have been proposed in the literature and achieved excellent performance. However, they are often fully supervised, thus requiring abundant pixel-level labeled samples, which is time-consuming and labor-intensive. Especially compared to the common single-temporal interpretation, labeling bi-temporal images is often more complicated. Therefore, this study combines weakly supervised learning (WSL) to reduce label acquisition costs. But changed regions are small, fragmented, and similar to the background, which increase the gap between weakly supervised and fully supervised tasks. To address these difficulties, we explore self-supervised methods to construct a WSL framework based on image-level labels for general CD, termed WSLCD in this paper. First, we design a double-branch siamese network to derive embeddings and initial class attention maps (CAMs), which inputs the original image pair and the spatially transformed image pair. Second, mutual learning and equivariant regularization (MLER) is enforced on CAMs from different views, which implements consistency constraints in confusion regions and makes CAMs learn from each other based on saliency regions. Furthermore, prototype-based contrastive learning (PCL) is designed such that unreliable pixels can learn from prototypes computed from reliable pixels. PCL includes intra-view contrast and cross-view contrast depending on whether the prototypes and class embeddings are from the same view. With the above strategies, we narrow the gap between image-level weakly supervised CD and fully supervised CD. Experiments are conducted on three CD datasets, including CLCD, DSIFN and GCD. Our method achieves state-of-the-art performance on pseudo label generation and CD. The code is available at https://github.com/mfzhao1998/WSLCD.

With the development of high-resolution satellites, more and more attention has been paid to remote sensing (RS) scene classification. Convolutional neural networks (CNNs), which replace the traditional handcrafted features with a learning-based feature extraction mechanism, are widely used in scene classification. But CNNs are less effective in deriving long-range contextual relations, which limits the further improvement. Visual transformer (VT), an emerging image processing method, provides a new perspective for RS scene classification by directly acquiring long-range features. Although there have been limited works combining CNN and VT through simple concatenation, the collaborations between them are insufficient. To address these issues, we propose a local and long-range collaborative framework (L2RCF). First, we design a dual-stream structure to extract the local and long-range features. Second, a cross-feature calibration (CFC) module is designed for them to improve representation of the fusion features. Then, combining deep supervision (DS) and deep mutual learning (DML), a novel joint loss is proposed to enhance the dual-stream feature extractor and further improve the fused features. Finally, a two-stage semi-supervised training strategy is designed to improve performance with unlabeled samples. To demonstrate the effectiveness of L2RCF, we conducted experiments on three widely used RS scene classification data sets: RSSCN7, AID, and NWPU. The results show that L2RCF performs significantly better compared with some state-of-the-art scene classification methods.

Remote sensing (RS) scene classification has attracted extensive attention due to its large number of applications. Recently, convolutional neural network (CNN) methods have shown impressive ability of feature learning in the RS scene classification. However, the performance is still limited by large-scale variance and complex background. To address these problems, we present a multilayer feature fusion network with spatial attention and gated mechanism (MLF2Net_SAGM) for RS scene classification. First, the backbone is employed to extract multilayer convolutional features. Then, a residual spatial attention module (RSAM) is proposed to enhance discriminative regions of the multilayer feature maps, and the key areas can be harvested. Finally, the multilayer spatial calibration features are fused to form the final feature map, and a gated fusion module (GFM) is designed to eliminate feature redundancy and mutual exclusion (FRME). To verify the effectiveness of the proposed method, we conduct comparative experiments based on three widely used RS image scene classification benchmarks. The results show that the direct fusion of multilayer features via element-wise addition leads to FRME, whereas our method fuses multilayer features more effectively and improves the performance of scene classification.

The mechanism of connecting multimodal signals through self-attention operation is a key factor in the success of multimodal Transformer networks in remote sensing data fusion tasks. However, traditional approaches assume access to all modalities during both training and inference, which can lead to severe degradation when dealing with modal-incomplete inputs in downstream applications. To address this limitation, we propose a novel approach to incomplete multimodal learning in the context of remote sensing data fusion and the multimodal Transformer. This approach can be used in both supervised and self-supervised pretraining paradigms. It leverages the additional learned fusion tokens in combination with modality attention and masked self-attention mechanisms to collect multimodal signals in a multimodal Transformer. The proposed approach employs reconstruction and contrastive loss to facilitate fusion in pretraining, while allowing for random modality combinations as inputs in network training. Experimental results show that the proposed method delivers state-of-the-art performance on two multimodal datasets for tasks, such as building instance/semantic segmentation and land-cover mapping when dealing with incomplete inputs during inference.

Semantic segmentation for remote sensing images (RSIs) plays an important role in many applications, such as urban planning, environmental protection, agricultural valuation, and military reconnaissance. With the boom in remote sensing technology, numerous RSIs are generated; this is difficult for current complex networks to handle. Efficient networks are the key to solving this challenge. Many previous works aimed at designing lightweight networks or utilizing pruning and knowledge distillation methods to obtain efficient networks, but these methods inevitably reduce the ability of the resulting models to characterize spatial and semantic features. We propose an effective deep supervision-based simple attention network (DSANet) with spatial and semantic enhancement losses to handle these problems. In the network, (1) a lightweight architecture is used as the backbone; (2) deep supervision modules with improved multiscale spatial detail (MSD) and hierarchical semantic enhancement (HSE) losses synergistically strengthen the obtained feature representations; and (3) a simple embedding attention module (EAM) with linear complexity performs long-range relationship modeling. Experiments conducted on two public RSI datasets (the ISPRS Potsdam dataset and Vaihingen dataset) exhibit the substantial advantages of the proposed approach. Our method achieves 79.19% mean intersection over union (mIoU) on the ISPRS Potsdam test set and 72.26% mIoU on the Vaihingen test set with speeds of 470.07 FPS on 512 × 512 images and 5.46 FPS on 6000 × 6000 images using an RTX 3090 GPU.