General-purpose foundation models have led to recent breakthroughs in artificial intelligence. In remote sensing, self-supervised learning (SSL) and Masked Image Modeling (MIM) have been adopted to build foundation models. However, these models primarily learn low-level features and require annotated data for fine-tuning. Moreover, they are inapplicable for retrieval and zero-shot applications due to the lack of language understanding. To address these limitations, we propose RemoteCLIP, the first vision-language foundation model for remote sensing that aims to learn robust visual features with rich semantics and aligned text embeddings for seamless downstream application. To address the scarcity of pre-training data, we leverage data scaling which converts heterogeneous annotations into a unified image-caption data format based on Box-to-Caption (B2C) and Mask-to-Box (M2B) conversion. By further incorporating UAV imagery, we produce a 12 × larger pretraining dataset than the combination of all available datasets. RemoteCLIP can be applied to a variety of downstream tasks, including zero-shot image classification, linear probing, k -NN classification, few-shot classification, image-text retrieval, and object counting in remote sensing images. Evaluation on 16 datasets, including a newly introduced RemoteCount benchmark to test the object counting ability, shows that RemoteCLIP consistently outperforms baseline foundation models across different model scales. Impressively, RemoteCLIP beats the state-of-the-art method by 9.14% mean recall on the RSITMD dataset and 8.92% on the RSICD dataset. For zero-shot classification, our RemoteCLIP outperforms the CLIP baseline by up to 6.39% average accuracy on 12 downstream datasets.

Of late, there are many studies on the robust discriminant analysis, which adopt L₁-norm as the distance metric, but their results are not robust enough to gain universal acceptance. To overcome this problem, the authors of this article present a nonpeaked discriminant analysis (NPDA) technique, in which cutting L₁-norm is adopted as the distance metric. As this kind of norm can better eliminate heavy outliers in learning models, the proposed algorithm is expected to be stronger in performing feature extraction tasks for data representation than the existing robust discriminant analysis techniques, which are based on the L₁-norm distance metric. The authors also present a comprehensive analysis to show that cutting L₁-norm distance can be computed equally well, using the difference between two special convex functions. Against this background, an efficient iterative algorithm is designed for the optimization of the proposed objective. Theoretical proofs on the convergence of the algorithm are also presented. Theoretical insights and effectiveness of the proposed method are validated by experimental tests on several real data sets.

The effective combination of hyperspectral image (HSI) and light detection and ranging (LiDAR) data can be utilized for land cover classification. Recently, deep learning-based classification methods, especially those utilizing Transformer networks, have achieved remarkable success. However, deep learning classification methods for multi-source data still encounter various technical challenges, such as the comprehensive utilization of multi-scale information, the lightweight network design, and the efficient fusion strategies for heterogeneous data. To address these challenges, we propose a novel and efficient deep neural network, namely multi-scale 3D-2D mixed CNN feature extraction and multi-source data lightweight attention-free fusion network (M2FNet) based on CNN and Transformer. Through end-to-end training, this network effectively combines heterogeneous information from multiple sources, leading to improved performance in joint classification. Specifically, M2FNet employs a multi-scale 3D-2D mixed CNN design to extract both the spatial-spectral features of HSI and the depth-based elevation features of LiDAR data. Subsequently, the extracted features are fed into a novel encoder comprising a feature enhancement module, designed with mathematical morphology and a dilated convolutional module derived from the self-attention of the conventional Transformer encoder (DConvformer), which plays a crucial role in integrating multi-source information within the network. The well-designed architecture enables the network to acquire multi-scale depth and high-order features, significantly reducing the number of training parameters. Comparative experimental results and ablation studies demonstrate that M2FNet outperforms other advanced methods. The source code is publicly available at https://github.com/cupid6868/M2FNet.git.

Recently, there are many works on discriminant analysis, which promote the robustness of models against outliers by using L₁- or L2,1-norm as the distance metric. However, both of their robustness and discriminant power are limited. In this article, we present a new robust discriminant subspace (RDS) learning method for feature extraction, with an objective function formulated in a different form. To guarantee the subspace to be robust and discriminative, we measure the within-class distances based on L2,s-norm and use L2,p-norm to measure the between-class distances. This also makes our method include rotational invariance. Since the proposed model involves both L2,p-norm maximization and L2,s-norm minimization, it is very challenging to solve. To address this problem, we present an efficient nongreedy iterative algorithm. Besides, motivated by trace ratio criterion, a mechanism of automatically balancing the contributions of different terms in our objective is found. RDS is very flexible, as it can be extended to other existing feature extraction techniques. An in-depth theoretical analysis of the algorithm's convergence is presented in this article. Experiments are conducted on several typical databases for image classification, and the promising results indicate the effectiveness of RDS.

Since remote sensing images of post-fire vegetation are characterized by high resolution, multiple interferences, and high similarities between the background and the target area, it is difficult for existing methods to detect and segment the burned area in these images with sufficient speed and accuracy. In this article, we apply salient object detection (SOD) to burned area segmentation (BAS), the first time this has been done, and propose an efficient burned area segmentation network (BASNet) to improve the performance of unmanned aerial vehicle (UAV) high-resolution image segmentation. BASNet comprises positioning module and refinement module. The positioning module efficiently extracts high-level semantic features and general contextual information via global average pooling layer and convolutional block (CB) to determine the coarse location of the salient region. The refinement module adopts the CB attention module to effectively discriminate the spatial location of objects. In addition, to effectively combine edge information with spatial location information in the lower layer of the network and the high-level semantic information in the deeper layer, we design the residual fusion module to perform feature fusion by level to obtain the prediction results of the network. Extensive experiments on two UAV datasets collected from Chongli in China and Andong in South Korea, demonstrate that our proposed BASNet significantly outperforms the state-of-the-art SOD methods quantitatively and qualitatively. BASNet also achieves a promising prediction speed for processing high-resolution UAV images, thus providing wide-ranging applicability in post-disaster monitoring and management.

The self-attention mechanism has been empirically shown its effectiveness in a wide range of computer vision applications. However, it is usually criticized for the expensive computation cost. Although some revised methods are proposed in the recent past, they are not maturely applicable to remote sensing scene (RSS) images. To address this problem, in this article, we propose a simple yet effective context acquisition module, named thrifty attention, which can capture the long-range dependence efficiently and effectively. Moreover, a recurrent version for thrifty attention, termed recurrent thrifty attention (RTA), is further proposed to take the long-range multihop communications in space–time for RSS images. RTA is a general global contextual information acquisition module that can be used in any hierarchy of deep convolutional neural networks. To demonstrate its superiority, we deploy it to the classical ResNet and establish our proposed RTA Network (RTANet). Extensive experiments are carried out on two levels of the RSS recognition tasks, i.e., the image-level RSS classification and the instance-level RSS object detection. Compared with the standard self-attention mechanism, RTA can reduce at most 0.43 M model parameters while increasing a slight of model floating-point operations per second (FLOPs). Furthermore, results on RSS classification and object detection further verify the accuracy superiority of RTANet.