Synthetic aperture radar (SAR) ship detection methods have achieved remarkable progress in recent years. However, unlike RGB images, the characteristics of SAR imaging will result in non-intuitive ...feature representations. Furthermore, due to the insufficient data of SAR images, existing methods relying on plenty of labeled SAR images may be hard to achieve promising performance. To address the aforementioned issues, a cross-modality feature transfer (CMFT) method is proposed in this article, which enhances feature representations in the SAR modality by transferring rich knowledge in the RGB modality. First, we propose a multilevel modality alignment network (MMAN), which encourages the model to effectively learn modality-invariant features and alleviate the large cross-modality discrepancies by aligning features from multilevels (scene level, local level, global level, and instance level). Second, to address the underperformance of samples with non-intuitive features in the modality alignment, we introduce a hard-sample supervision module (HSM) in the stage of feature extraction, which can thoroughly exploit the feature of hard-to-align samples by giving more optimization energy for them. Third, to enhance the discriminability of instance-level features, a feature complementary module (FCM) is customized to fully explore the potential complementary clues between instance-level features and context information for the instance-level feature alignment. Extensive experimental results demonstrate that the CMFT outperforms the state-of-the-art detectors. Compared to the baseline model, CMFT improves the accuracy by 3.1% mean average precision (mAP) on the SSDD dataset and 3.4% mAP on the HRSID dataset, demonstrating its superior SAR ship detection performance.
Rapid and reliable gully erosion (GE) extraction from high-resolution remote sensing (HRRS) images is crucial for the development of land protection measures. For this task, semantic segmentation ...methods are widely considered the state-of-the-art solutions. Nevertheless, providing sufficient and clean training labels for segmentation models demands substantial expenses and time investment. In this context, a label correction learning (LCL) framework is proposed to effectively extract GE from HRRS images by leveraging more readily available noisy labels. The core objective of this framework is to suppress the adverse impact of noise within noisy labels on the model performance. To achieve this, we introduce three key components in the framework, including an adaptive correction loss function, a multitree refinement module, and a noise correction module. These components collaborate to rectify noisy labels during training, thereby providing the model with a training set containing less noise. To validate the effectiveness of the LCL framework, three severely eroded regions in Northeast China are selected as study areas and corresponding noisy datasets are generated. Extensive experiments on these datasets demonstrate that our framework can significantly mitigate the negative influence of label noise and ultimately achieve superior GE extraction performance. Moreover, by employing the proposed framework, we generate GE coverage maps for the study areas and obtain measurements of gully area and length that are very close to the true statistics. Such a framework that can effectively learn from noisy labels holds promise as a practical and cost-efficient means to provide reliable data references for land resource protection.
With the development of Earth observation technology, it becomes easier and easier to acquire multi-modal image data at the same time. To improve the performance of multi-modal remote sensing ...detection algorithm, a new fusion feature optimization detection network (FFODNet) is proposed. The method is designed to solve the problem of performance degradation caused by the unreliability of single modal data in multi-modal remote sensing data. The key to obtain high quality fusion features from multi-modal data with interference is to suppress single modal redundant features and fully integrate multi-modal features. The proposed method mainly includes two improvements. Firstly, a novel joint expression optimization module (JEOM) is designed to enhance the target features and suppress the redundant and interference features that affect the fusion effect. Additionally, we propose a novel specific information enhancement module (SIEM) to further enhance the discriminative feature information of targets within each modal image. Experiments on DroneVehicle dataset show that our proposed method is state-of-the-art on this dataset.
China has just started the study of undersea feature naming, compared to the USA, Russia, Germany and Japan, as rules of undersea feature naming and building of techniques and methods have not been ...systemized. Proposing names for the newly-discovered undersea features and participating in setting standardization of undersea feature names will help to enhance China's international status and influence, and also show China's international obligation and state strength.
1 Change detection has always been a hot research area in the field of hyperspectral image (HSI) processing. However, in the current change detection methods, most of them need to train a large ...number of labeled data to extract representative features. In this paper, a hyperspectral change detection method based on multi-scale three-dimensional (3D) convolution autoencoder network (M3CAN) is proposed. Firstly, the multi-scale 3D convolution block is adopted in the autoencoder which can extract effective spectral-spatial joint features of HSIs. Then, the autoencoder is pre-trained to obtain the trained encoder as the feature extractor. Finally, the feature maps of the bi-temporal data are obtained by the encoder and then sent to the Softmax classifier to obtain the final change detection result. In this paper, unsupervised training of autoencoder is combined with supervised training of classifier. Therefore, only a small amount of data is needed to complete the training, which avoids the difficulty of requiring many labeled training data. Experiments show that the proposed method has good results on two datasets.
Nowadays, graph convolution networks are getting more and more attention in the field of hyperspectral image classification. The graph convolution can be divided into long-range and short-range graph ...convolution (GConv). However, the two graph convolutions cannot acquire global and local features at the same time, making the node features may not be accurate enough. Therefore, we propose a novel graph convolution approach, called short and long range graph convolution (SLGConv), which combines the advantages of long-range and short-range GConv. SLGConv can extract long-range (global) and short-range (local) spatial-spectral features, eliminating the disadvantages of each of long-range and short-range graph convolution. Furthermore, SLGConv can ensure that the features of nodes are not smoothed in the convolution process. Then, three layers of SLGConv are used to form the short and long range graph convolution network (SLGCN) for hyperspectral image classification. Experiments on three HSI datasets indicate that the SLGCN can obtain better classification performance when compared with seven state-of-the-art methods.
Without any prior information of anomalies or background, hyperspectral anomaly detection has received a wide attention. However, such unsupervised style brings difficulties in training and learning ...effective features of hyperspectral image to perform detection. This paper proposes a novel hyperspectral anomaly detection algorithm using bilateral-filtered generative adversarial networks (BFGAN). Bilateral filter can smooth images and remove anomalous points while preserving edges. With closeness weights and similarity weights, the bilateral-filtered hyperspectral image can be considered as background data, so that hyperspectral background labels are obtained. Only with one class of labels, the structure of generative adversarial networks has an ability to solve two-class problem. By using the filtered background data and their labels, generative adversarial networks are trained to improve discriminator's discriminative capability for background data in a competing style. Finally, the model discriminator can finally output big probabilities for background samples and small probabilities for anomalous samples. Experiments on two real hyperspectral images demonstrate that the proposed method outperforms other state-of-the-art competitors.
Existing deep learning-based models for hyperspectral image classification (HSIC) may be suboptimal in the utilization and the balance between the spatial and spectral information, and they pay more ...attention to the design of architecture and modules but ignore the generalization of the model on the data with slight distribution shift. In this paper, a novel framework based on spectral-spatial convolutional Transformer and Mixup regularization (SSCT-M) is proposed for hyperspectral image classification. SSCT-M designs a dual-branch architecture to extract spectral and spatial features and adopts a Transformer-based network to adaptively balance the contribution between spectral and spatial pipelines. Then, SSCT-M adopts a two-stage mixup regularization to supervise the learning of the model, which can boost the generalization performance of the model by learning convex combinations on the labeled data while encouraging the model to make consistent predictions on the perturbed unsupervised data. Extensive experimental results on two HSI benchmarks demonstrate the effectiveness of our proposed framework.
In recent years, deep learning has demonstrated its transformative potential in the field of hyperspectral image (HSI) processing but is notoriously data-hungry. However, wanting to obtain a large ...number of labels is labor-intensive and time-consuming. To reduce the dependence of the model on the label samples while maintaining high detection accuracy, a hyperspectral image change detection algorithm based on active learning strategy (ALCD) is proposed. First, the active learning strategy is employed to select high-value labeled samples from the test set as additional training data, gradually enhancing the model's detection performance. Second, the self-attention module MOAT is introduced to enable effective interaction of local information during the feature extraction process and enhance the network's feature expression capability. Then, the feature interaction and the mixing block are used to blend the features of the bitemporal images, so that the feature distribution of the bitemporal images is more similar, which is conducive to subsequent feature extraction and classification. Experiments on two HIS datasets show that the proposed method can obtain better change detection results than the four comparison algorithms.
Erosion gullies are a prominent manifestation of soil erosion. And timely and accurate acquisition of relevant data about erosion gullies plays a crucial role in their management and control. ...Currently, there is a deficiency in automation within the majority of erosion gully detection methods. The post-classification comparison method using semantic segmentation techniques and the direct change detection method often struggle to ensure high accuracy. Therefore, a end to end change detection method based on deep supervised and feature interaction (DSFNet) is proposed for erosion gullies in this paper. To achieve accurate localization of erosion gully semantic information, DSFNet employs a deep supervision strategy to constrain the semantics of erosion gullies. Furthermore, in order to extract representative features related to erosion gullies and improve the detection accuracy of the model, a feature interaction and upsampling module (IUModule) is employed. Experimental results show that DSFNet exhibits better performance on erosion gully dataset.
