Utilizing simulated P-band PolInSAR backscatter data in pine forests, we qualitatively and quantitatively analyzed the applicability of the RVoG model in tree height inversion. Small radar incidence angles and low forest stand densities can decrease canopy continuity, invalidating a critical RVoG model assumption. Also, the ground backscatter decreases at a large radar incidence angle, reducing the accuracy of ground phase estimation. A high stand density can ensure good canopy continuity and stable tree height inversion. Thus, for better inversion, the radar incidence angle should be around 35°, and the stand density should be ≥ 300 stems/ha.

The impact of cross-polarized (i.e.,${S_{HH}}{S^{\ast}}_{HV},{S_{VV}}{S^{\ast}}_{HV},{\text{ }}{S_{HV}}{S^{\ast}}_{HH},{\text{ }}and{\text{ }}{S_{HV}}{S^{\ast}}_{VV}$) backscatter on analyzing scattering mechanisms in PolSAR datasets of radar targets from forested and oriented urban areas was studied. A new decomposition algorithm was proposed by considering all nine elements in the covariance matrix. It was eigenvalue/eigenvector-based. Its validity and effectiveness in separating radar targets from forested and oriented urban areas were assessed with the Radarsat-2 C-band PolSAR backscatter data. The study area is San Francisco, California, USA. Satisfactory separation results are achieved. Compared with the three existing decomposition algorithms, the proposed algorithm should be most capable of distinguishing radar targets from forested and oriented urban areas and help improve the forest and urban parameter inversion using PolSAR backscatter data.

Sea surface temperature (SST) is one of the most important parameters in the global ocean-atmosphere system. Predicting SST can help to analyze and identify extreme weather and protect marine environment in advance. Traditional numerical and machine learning methods tend to ignore spatial features. The single model in existing deep learning methods suffers from weakening spatial features and reducing ability of discriminating time-series information. At the same time, rare consideration about the influence of ocean physical phenomena has been given. These will lead to inaccurate prediction results. Based on the spatial-temporal characteristics and physical laws of the SST field, this paper proposes a hybrid multi-scale spatial-temporal model combining error correction map (ECM-HMSTM) to predict the SST. First, the ECM-HMSTM can comprehensively extract the spatial-temporal features of the SST field at different scales and thus the SST prediction map can be obtained. Second, by a new error correction approach based on the activity of tropical instability waves (TIWs), the ECM-HMSTM can effectively predict the variation characteristics of TIWs signals, which results in producing the error correction maps. Third, by fusing the two above maps, the SST field in the tropical eastern Pacific Ocean after five days is predicted. Experiment results show that the accuracy of the ECM-HMSTM was improved by 10.3% compared with the current state-of-the-art deep convolution model. Moreover, the SST predicted by the ECM-HMSTM performs well on characterizing the intensity of TIWs. Therefore, this paper provides a strategy for effective short-term prediction of SST fields, which is of guidance for prediction and analysis of ocean phenomena and climate.

The detection of ships encompasses an abundance of applications within the domains of fishery management, marine rescue operations, and maritime monitoring. In recent years, a multitude of detectors based on deep learning have been utilized for the purpose of ship detection using synthetic aperture radar (SAR) images. However, disturbed by the strong scattering background on land and the influence of the SAR target scale, the existing detectors face great challenges in detecting inshore small ships. To solve this problem, this article proposes an oriented ship detection network for SAR images based on soft threshold and context information. First, a soft-threshold quantization module (STQM) based on the soft threshold function is designed to alleviate the interference of background noise on the feature map. Second, a local and global context fusion module (LGCFM) is designed to capture the contextual information of the target to enhance the detection of small targets. Third, the inclusion of the center loss in the loss function serves to impose additional constraints on the center coordinates and shape of the oriented bounding box. This is done to achieve a more balanced distribution of loss contributions across the various variables and to mitigate the target’s susceptibility to variations in the shape of the ground-truth bounding box. Finally, the proposed network is tested on the publicly available mini-rotated-high-resolution SAR images dataset (MR-HRSID) and Rotated SAR Ship Detection Dataset (R-SSDD) datasets. The results from experiments demonstrate that our approach attains state-of-the-art detection capabilities for inshore ships and small targets, while also effectively mitigating interference from terrestrial noises.

Thanks to the powerful feature extraction and expression ability of convolutional neural networks (CNNs), exceptional success has been achieved in the field of ship automatic target recognition (ATR) of synthetic aperture radar (SAR). However, the CNNs cannot work effectively with sparse labelled samples and imbalanced categories.This study proposes a new Attention-Dense-CycleGAN (ADCG) method that is suitable for the ship transfer learning task from optical to SAR (OPT2SAR). The key improvement of the ADCG lies in the construction of a Dense Connection Module (DCM) and a lightweight Convolutional Block Attention Module (CBAM). The DCM is able to overcome the problems of generator feature redundancy, large network model parameters, and severe training time in the original CycleGAN network. The lightweight CBAM can solve the problem of not being able to locate the main features of ships with a minimal increase in network parameters. Compared with the performance of other popular generative adversarial networks, the superior performance of the ADCG in the OPT2SAR transfer learning is demonstrated with the Fréchet Inception Distance (FID) minimum of 76.04 and the Kernel Inception Distance (KID) minimum of 0.0403. Finally, the ability of pseudo-SAR domain images were tested to improve the recognition accuracy of popular ship classification networks, this achieved an average improvement of 6% in recognition accuracy. Therefore the results of this study verifies the rationality, validity, and application value of pseudo-SAR domain in solving the problems of sparse marker samples and class imbalance in ship ATR network model.

Fine-grained ship classification in optical remote sensing images is a major challenge in the ocean observation field, elaborated as follows: First, the cost of acquiring ship images is expensive. Obtaining numerous labeled samples is difficult, resulting in the poor generalization ability of training models. Second, the features of ship target cannot be accurately obtained owing to complex background interference. Third, inter-class similarity and intra-class diversity among different ships render ship classification difficult. In this study, we propose LATA-BP-BDC: a bi-prototype Brownian Distance Covariance (BDC) metric network with lightweight adaptive task attention (LATA) for few-shot fine-grained ship classification. First, the LATA module is used to generate 3-dimensional (3D) weights, which can effectively reduce complex background interference and improve the adaptive capturing ability of target features without including additional network operators. Second, we input target features into the BDC metric module and output the BDC matrices to represent image information. Because the similarity between two images can be calculated as the corresponding BDC matrices distance, the improvement of the relevance of similar targets can be realized. Finally, we use the bi-prototype module to generate highly accurate prototypes, further calibrating information differences between images, which enhances the correlation between the same category samples and separability between different categories samples. Consequently, this process effectively reduces the influence of large intra-class appearance variation and small inter-class appearance variation. We perform validation using two fine-grained datasets, FGSCR and CUB. Compared with state-of-the-art methods, the LATA-BP-BDC achieves a superior performance and has good generalization for fine-grained few-shot classification.

Model-based decomposition methods are widely used in full-polarization synthetic aperture radar (SAR), for the inversion and interpretation of ground features and constitute an important approach for understanding the behavior of backscattering. However, owing to the substantial differences between land and marine environments, different man-made and natural vegetation scattering structures render existing decomposition models unable to reasonably characterize scatterers on ships. Moreover, the combination of polarization decomposition models and neural networks for ship detection has rarely been investigated. Therefore, this study proposes a four-component decomposition model (Ship-4SD) suitable for describing the scattering characteristics of ships based on the surface scattering, double-bounce scattering, ±45° oriented dipole, and asymmetric scattering components. Furthermore, based on the differences in the scattering properties exhibited by different scattering components in ships and the powerful feature extraction capability of convolutional neural networks (CNNs), a scattering characteristic-aware fully polarized SAR ship detection network (SCANet) was designed to make full use of the scattering components in the decomposition model. Finally, the experimental results on a large amount of GF-3 fully polarized SAR data validated that the reasonability and superiority of Ship-4SD and SCANet. The Ship-4SD can better distinguish ship and clutter pixels compared to other four-component models and has a higher target-clutter ratio with respect to the multi-component models. SCANet proposed in this paper achieved an average precision of 94.43% and 96.56% on the GF-3 and SSDD datasets, respectively, which is better than that of other competitive CNN algorithms.

A two-step algorithm to delineate urban targets in PolSAR data was studied. First, we used an eigenvalue- and eigenvector-based decomposition algorithm without the azimuth symmetry assumption to extract three types of scattering components of radar targets. The odd scattering of urban targets was weaker than that of non-urban targets. The cross scattering in urban areas with small azimuth orientation (AO) angles was smaller than that in urban areas with large AO angles. Then, considering the relative importance of the odd and cross scatterings, urban targets were delineated. The algorithm was verified using PALSAR2 L-band data, San Francisco, California (CA), USA, where four types of radar targets, including urban targets with small and large AO angles, water surface, and vegetated areas, were studied. Urban targets of a wide range of AO angles have been effectively identified.

Trees in flat and wet forest floors and urban targets with small orientation angles can produce strong double scattering. Then, misclassification of trees as urban targets occurs in the PolSAR data decomposition. This study proposed an improved decomposition algorithm for the urban and forest targets classification. First, strong double scattering between forest and urban targets was differentiated, and then forest targets were treated as azimuthally symmetric radar targets in the PolSAR decomposition algorithm. The improved decomposition algorithm was verified by two L-band PolSAR datasets, a UAVSAR dataset northwest New Bern, NC, USA, and one ALOS PALSAR dataset San Francisco, CA, USA. In the datasets, the double scattering was dominant for forested areas. The double scattering dominated some urban areas, but other urban areas had strong volume scattering. Nevertheless, the revised decomposition algorithm correctly delineated both types of targets and extended the usability of the original decomposition algorithm.

Urban targets with large azimuth orientation angles and vegetation targets have strong volume scattering in polarimetric synthetic aperture radar (PolSAR) data. Their separation is a challenge. The real parts of ShhShv* and SvvSvh* correlate to the azimuthal symmetry of radar targets in the PolSAR data. The values are close to zero for azimuthally symmetric vegetation targets and much away from zero for azimuthally asymmetric urban targets. Also, in the urban area with large azimuth orientation angles, the real part values of ShhShv* are negative, but the real part values of SvvSvh* positive. Thus, the sum of the real parts of ShhShv* and -SvvSvh* are studied as a new separation descriptor, Sd, to delineate the urban targets from vegetation targets. The descriptor is evaluated by using the 2009 PALSAR San Francisco PolSAR data. The Sdvalues are negative in the urban area where targets with large azimuth orientation angles exist but close to zero in the vegetated area. Correct separation rates between the urban and vegetation targets are 92.44%-99.5%, with an empirical threshold. Thus, the descriptor can effectively differentiate urban targets with large azimuth orientation angles from vegetation targets.

With a moving SAR platform, the HV (H transmitted and V received) and VH (V transmitted and H received) components in the polarimetric synthetic aperture radar (PolSAR) data are obtained at different positions along the azimuth direction. The HV and VH datasets' correlation coefficient varies whether a ground radar target is stationary or not. A building is standing, but a tree canopy may not be due to the surrounding air movement. Therefore, based on the correlation between the HV and VH components in one single PolSAR dataset, a stationary descriptor,$S_{D}$, was studied to separate the volume scattering from a stationary urban target with a large azimuth orientation angle or a non-stationary vegetation canopy. The separability was evaluated using two PALSAR PolSAR datasets. The correct separation rates were 86.2% for urban targets with large azimuthal orientation angles and 93.9% or higher for vegetation targets. Thus,$S_{D}$effectively differentiates urban targets with large azimuth orientation angles from vegetation targets.

The tree height estimation using the simplified three-stage algorithm and HH and HV polarizations or dual-polarization data was studied. First, we considered (HH+2×HV) as VV data obtaining the simulated quad-polarization data. Then, the polarization interference coefficients were computed. To reduce the combined effort of the HH and HV polarizations on the interference coefficients, we revised the coefficients. Both sets of coefficients were used in the estimation of tree heights. The simulated mean tree height was 18 m. The estimation was improved when the revised coefficients were used. The mode of heights changed from 20.0 m to 18.4 m before and after the revision. Thus, the estimated tree heights should be acceptable. The required input of the quad-pol data to the three-stage algorithm was simplified with the input of the dual-pol data.

After revaluating the validity of the assumption made in the existing decomposition algorithms for the quad-pol SAR datasets, we proposed a two-step algorithm to resolve the invalid assumption. The algorithm was assessed using the quad-pol ALOS/PALSAR and NASA/JPL UAVSAR datasets. The results were satisfactory. The algorithm should be valid.

An algorithm to remove the impact of man-made targets on tree height estimation using the three-stage algorithm and PolInSAR data was studied. The data were the German/DLR E-SAR L-band PolInSAR data near Oberpfaffenhofen, Germany. Within the data, there were forested areas and man-made targets such as runways and buildings with variable azimuth angles. The runways with smooth surface or buildings with large azimuth angles were sequentially removed by the mask of surface and non-surface scattering components derived from the PolSAR decomposition algorithm, and the mask of high and low coherence coefficient. Then, tree heights were estimated. Results were satisfactory. Therefore, the developed algorithm was valid, and the applicability of the three-stage algorithm has been extended.

A new four component decomposition algorithm for urban targets delineation from PolSAR imagery was studied. First, a new correlation coefficient was introduced to describe characteristics of urban targets. The coefficient was the linear combination of two elements of a coherent matrix. Then, the coefficient was used to modify the volumetric scattering model component in Yamaguchi et al.'s four-component decomposition algorithm with the azimuth deorientation. Thus, a new four-component decomposition algorithm was established. With the algorithm, the percentage of double-bounced scattering increased but the percentage of volumetric scattering decreased in urban areas. Thus, the urban targets whose azimuth angles are 0° or non-zero° were effectively identified.