Band selection is an effective method to reduce redundancy in a hyperspectral image (HSI) without compromising the original contents. Popular band selection methods usually use strong assumptions, such as linear or nonlinear assumptions with simple predefined Kernel functions, to model the correlations between bands. However, this kind of strong assumption may not valid in the real environment due to the complex interactions between bands. In this letter, we treat hyperspectral band selection as a spectral reconstruction task. By assuming that an HSI can be sparsely reconstructed from a few informative bands, we propose an attention-based autoencoder to model the underlying nonlinear interdependencies between bands. The proposed model consists of two parts: an attention module and an autoencoder. The attention module is used to produce the attention mask which selects the most informative bands for every pixel. The autoencoder uses these informative bands to reconstruct the raw HSI. The final band selection is conducted via clustering column vectors of the attention mask and exploring the most representative band for each cluster. Different from most of the existing band selection methods, the proposed method directly learns global nonlinear correlations between bands without strong assumptions. The proposed model is easy to implement and all the parameters can be jointly optimized using the stochastic gradient descend algorithm. Experiments on three open public data sets show that the proposed method offers the promising results.
31 July 2023 in IEEE Transactions on Geoscience and Remote Sensing
Object detection in remote sensing is a challenging task due to the arbitrary orientations of objects and the vast variation in the number of objects within a single image. For instance, one image may contain hundreds of small vehicles, while another may only have a single football field. Recently, DEtection TRansformer (DETR) and its variants have achieved great success in object detection by setting a fixed number of object queries and using bipartite graph matching for one-to-one label assignment. However, we have observed that bipartite graph matching can result in relative redundancy of object queries when the number of objects changes dramatically in an image. This relative redundancy can cause two problems: slower convergence during training and redundant bounding boxes during inference. To analyze the aforementioned problems, we proposed a metric, Redundancy of Object Query (ROQ), to quantitatively analyze the redundancy. Through experiments, we discovered that the reason for the two issues is the difficulty in distinguishing between high-quality negative samples and positive samples. In this paper, we proposed Efficient-Matching Oriented Object Detection with Transformers(EMO2-DETR) consisting of three dedicated components to address the aforementioned issues. Specifically, Reassign Bipartite Graph Matching (RBGM) is proposed to extract high-quality negative samples from the negative samples. And Ignored Sample Predicted Head (ISPH) is proposed to predict high-quality negative samples. Then, Reassigned Hungarian loss is used to better involve high-quality negative samples in the update of model parameters. Extensive experiments on DOTAv1 and DOTAv1.5 datasets demonstrated that our proposed method achieves competitive results.
12 August 2024 in International Journal of Applied Earth Observation and Geoinformation
Hyperspectral unmixing, which estimates end-members and their corresponding abundance fractions simultaneously, is an important task for hyperspectral applications. In this article, we propose a new autoencoder-based hyperspectral unmixing model with three novel components. First, we propose a new sparse prior to abundance maps. The proposed prior, called orthogonal sparse prior (OSP), is based on the observations that different abundance maps are close to orthogonal because, generally, no more than two end-members are mixed within one pixel. As opposed to the conventional norm-based sparse prior that assumes the abundance maps are independent, the proposed OSP explores the orthogonality between the abundance maps. Second, we propose the hyper-Laplacian loss to model the reconstruction error. The key observation is that the reconstruction error distribution usually has a heavy-tailed shape, which is better modeled by the hyper-Laplacian distribution rather than the commonly used Gaussian distribution. Third, to ease the side effect of outliers for end-member initializations, we develop a data-driven approach to detect outliers from the raw hyperspectral images. Extensive experiments on both synthetic and real-world data sets show that the proposed method significantly and consistently outperforms the compared state-of-the-art methods, with up to more than 50% improvements.
6 May 2024 in IEEE Geoscience and Remote Sensing Letters
The topic of object detection (OD) in remote sensing (RS) has received a lot of attention due to the rapid growth of deep learning. However, RS images typically have the following characteristics: significant variations in object scales, tight arrangement of small objects, and indistinguishable feature boundaries between objects and backgrounds. These challenges lead to defects, such as insufficient feature extraction and information loss of the existing methods. To address the above issues, based on the YOLOv7 architecture, we present a novel OD method named GDRS-you only look once (YOLO). Our primary contributions include: first, an enhanced feature extraction network based on deformable convolution is proposed to improve the network’s ability to model geometric transformations. Second, we abandoned the traditional feature pyramid architecture and construct a multiscale feature aggregation network based on the gather-and-distribute mechanism, which makes effective use of the feature obtained from the backbone and reduces the loss of information in the transmission process. Finally, the normalized Wasserstein distance (NWD) is introduced for hybrid loss training, which alleviates the sensitivity of the IoU-based metric to the location deviation of tiny objects. We demonstrate the effectiveness of GDRS-YOLO on the publicly available datasets NWPU VHR-10 and VisDrone datasets. Compared to the original YOLOv7, the proposed method improves the mean average precision (mAP) by 1.9% and 5.5%, respectively. These results emphasize the superior performance of the proposed model, which provides an efficient multiscale feature fusion solution for RS applications.
Satellite-based PM2.5 estimation is an effective means to achieve large-scale and long-term PM2.5 monitoring and investigation. Currently, most of methods retrieve PM2.5 from satellite-derived aerosol optical depth (AOD) or top-of-atmosphere reflectance (TOAR) during daytime. A few algorithms are also developed to retrieve nighttime PM2.5 from the satellite day–night band and the accuracy is greatly limited by moonlight and artificial light sources. In this study, we utilize the properties of absorption pollutants in infrared spectrum to estimate PM2.5 concentrations from satellite infrared data, thus achieve the PM2.5 estimation in both day and night. Himawari-8 infrared bands data are used for PM2.5 estimation by a specifically designed neural network and loss function. Quantitative results show the satellite derived PM2.5 concentrations correlates with ground-based data well with R2 of 0.79 and RMSE of 15.43 μg · m−3 for hourly PM2.5 estimation. Spatiotemporal distributions of model-estimated PM2.5 over China are also analyzed, and exhibit a highly consistent with ground-based measurements. Dust storms, heavy air pollution and fire smoke events are examined to further demonstrate the efficacy of our model. Our method not only circumvents the intermediate retrievals of AOD, but also enables consistent estimation of PM2.5 concentrations during daytime and nighttime in real-time monitoring.