This paper introduces a change detection method for Synthetic Aperture Radar (SAR) images acquired from different satellite orbits. Conventional methods struggle with angle sensitivity in SAR images, suffering from poor pixel correspondence between images. We propose an end-to-end framework that incorporates a semi-learnable feature transformation module. This module employs a re-sampling matrix, grounded in SAR imaging principles, to effectively align features extracted from images captured at disparate angles. Our approach is validated through three distinct scenarios: change detection in car parking lots across various orbits, also with orbital angles that are unseen during training, and flood detection where the training and testing using different satellite sources. The experimental results confirm that our method either achieve a competitive level equivalent to the other established methods, or effectively surpasses them in diverse scenarios.

This paper proposes a statistical anomaly detection method using persistent scatterer interferometry (PSI) . Change detection methods utilizing phase information in synthetic aperture radar (SAR) images are sensitive to various changes, including regular changes such as orbital changes, slow displacements, atmospheric effects, and observation noises, making it difficult to detect true anomalies. To address this issue, the proposed method employs the PSI results to construct a statistical model of regular changes, which can predict a new image. Anomalies are identified as pixels in the actual image that exhibit significant deviations from the prediction. The effectiveness of the proposed method is demonstrated through its application to real SAR data, where the potential damaged areas are easily discernible.

Semi-supervised domain adaptation offers a solution to the issues of data scarcity and annotation difficulty in Synthetic Aperture Radar (SAR) images by transferring knowledge from an existing, labeled source domain dataset to a minimally labeled target domain. However, current methods struggle with unique classes exclusive to the target domain and the inherent sampling bias towards specific acquisition angles present in the limited labeled target data, leading to learning an inadequate representation of the comprehensive target domain. This paper introduces a unified framework that leverages the class- and domain-independent SAR imaging physics to perform feature transformation across varying acquisition angles. The feature transformation module augments the sparsely labeled target data, thereby contributing to a more complete representation of the full target domain. Experimental results show that our proposed method outperforms existing techniques, highlighting the effectiveness of physics-guided feature extraction in bridging domain and class discrepancies amidst SAR images' angle sensitivity.

This paper proposes two techniques to improve the imaging quality of the SVD-Stolt algorithm when the synthetic aperture radar signal is acquired with high-squint and curved orbit geometry. The SVD-Stolt algorithm is known as an efficient imaging algorithm that works with any acquisition geometry using a numerical optimization on the simulated echo from reference points. This paper suggests two modifications to improve the image quality: setting the reference points along the beam direction and adding a new correction term. The experimental results show that error in the numerical optimization is reduced, and the imaging quality is improved when the squint angle is large and the orbit is curved.

This paper proposes a change detection method robust to orbital variation using Bayesian SAR tomography. The proposed method stochastically generates background images for each observation using Bayesian SAR tomography. These background images are then compared probabilistically with the observed SAR images to detect changes. This paper’s objective is to detect changes within the observation area by comparing time-series SAR images and identifying differences from the unchanged state in each image. However, the issue with this type of analysis is that it is difficult to generate a background image representing unchanged state due to orbital variation or the effects of random noise. The proposed method can generate background images adjusted to orbit positions for each observation using 3-dimensional information of the unchanged state reconstructed by Bayesian SAR tomography. The Bayesian approach allows us to evaluate the effect of random noise on each observation probabilistically. We apply the proposed method to SAR images obtained by TerraSAR-X satellite and find that it realizes reliable change detection in areas where it is difficult to detect changes using conventional methods.

This paper presents a semi-supervised domain adaptation method for SAR target recognition. The proposed method only requires a few real data to be labeled. The challenge is that due to the high angle sensitivity of SAR images, a network can easily overfit the training data at seen angles and fails to classify data at unseen angles. To overcome this, we design a conversion module that can infer what CNN features of images at unseen angles look like. This conversion module is designed as a rigid-body transformation followed by a conditional generative network. This design enables the network to gain high-level 3D understanding from 2D images. Thus, we name our network 3D converter. The network mainly learns from fully-labeled simulated SAR data and then the knowledge is adapted to fit the scarcely-labeled real SAR data. Our method improves the performance over the baseline by 3.48% on MSTAR benchmark when only 10 images/class are labeled. It also achieves comparable results to strongly supervised methods.

This paper presents a semi-supervised domain adaptation (SSDA) method for Synthetic Aperture Radar (SAR) image classification. SAR imagery is important in ground activity monitoring, but its wide application is impeded due to a lack of annotations. SSDA methods transfer class-discriminative knowledge from a fully-labeled source dataset to a scarcely-labeled target dataset. However, conventional methods often train models which overfit to labeled target data and fail on unlabeled data. To overcome this, we propose to additionally adapt intra-class variations. Specifically, a conversion network is trained to learn from source data the image feature variations caused by the change of image capturing angle. Then synthetic data, which represent a generalized target domain distribution, are estimated by applying the conversion to labeled target data. Our method improves the accuracy of the state-of-the-art SSDA approach from 64.28% to 80.40% in three-shot cases on the SAR ground vehicle dataset.

This paper proposes a method of inter-orbit change detection for high-resolution synthetic aperture radar (SAR) imagery using a conditional Siamese network. The proposed method introduces a sub-network with a condition, which is satellite orbit information, into the multitask Siamese network (MS-net). To introduce the conditions effectively, the weights sharing in one fully-connected layer after connecting the subnetwork is canceled. These tricks enable the proposed network to learn the absorption of layover effects depending on the orbit, which improves change detection performance. Experiments were conducted for detecting car changes in a parking lot by using Asnaro-2 images captured from five different orbits. Compared with the conventional MS-net, the proposed model improves AUC-ROC by 0.015 on average and is more robust to input orbit combinations.

This paper presents a small object, represented by approximately ten pixels in an image, change detection method based on multitask Siamese network for multitemporal SAR images. In our proposed method, not only change detection task but also object classification task is introduced to the network. The classification task is expected to enhance the performance of change detection by providing semantic information of changes and to focus attention of the network towards the target small object class. We tested the proposed method for a real-world application of car parking lot monitoring with 1-meter resolution TerraSAR-X images. Experimental results show that the f-measure of change class is improved by more than 7% over conventional methods based on post-classification, PCA+K-means and Siamese network. Furthermore, car-to-car type change is detected by the proposed method with 25% higher accuracy over the method without the classification task.

This paper proposes a method of cloud shadow removal for multispectral images to retrieve the ground reflectance of areas shadowed by clouds. Cloud shadows are cast when incident direct solar irradiance gets attenuated by clouds. To retrieve the ground reflectance of the shadowed pixels, it is required to estimate pixel-wise attenuation factor for the solar irradiance. Unlike conventional methods, the proposed technique takes the physical model of cloud shadow formation into account to accurately estimate the attenuation factors. According to the physical model, the factors are derived from the transmittance of an occluding cloud. Visual and quantitative results demonstrate that the proposed method outperforms the well-known de-shadowing algorithm. The average correlation coefficient of the corrected image with a reference image is improved from 0.45 to 0.75 by the proposed method as compared to the conventional method. Further, the average spectral angle with a reference image is improved by 10%.

A new high dynamic range rendering (HDRR) suitable for YUV images is presented. The method overcomes the problem on saturation decrease caused by the conventional HDRRs for YUV images. U and V components are corrected with an approximating function that is designed to preserve the metric chroma in CIELAB before and after an HDRR. The function is derived from the relationship between the changes in the Y component and those in the chroma. The color differences between the images corrected by our method and the ideal images were found to be in an acceptable range, where they can be recognized as being the same color. Our method achieved the highest score in a subjective assessment of preserving the perceptual chroma when it was compared with other methods.