Conference Agenda

Urbanization is continuing at an unprecedented rate in many cities across the globe. Timely and reliable information on the sprawl of settlements are important to support sustainable planning. Earth observation has been playing a crucial role to map land cover changes associated with urbanization (Ban & Yousif, 2016). Many studies have been conducted to demonstrate the potential of Synthetic Aperture Radar (SAR) and multispectral data for urban change detection (e.g. Ban & Yousif, 2012; Bovolo & Bruzzone, 2015; Bruzzone & Prieto, 2000; Gamba et al., 2006).
In recent years, several deep learning methods using fully Convolutional Neural Networks (CNNs) have been used to detect changes in multi-temporal satellite imagery. In particular, the vast amount of high resolution (10–20 m) imagery collected by the Sentinel-2 MultiSpectral Instrument (MSI) mission has been used extensively for urban change detection (e.g., Daudt et al., 2018; Papadomanolaki et al., 2021). More recently, we demonstrated that fusion of Sentinel-1 SAR and Sentinel-2 MSI data can improve urban change detection results (Hafner et al., 2021). While our dual stream network using late fusion achieved improvements over input-level fusion, recent work demonstrated that multi-modal deep neural networks suffer from their greedy nature, meaning that they tend to rely on just one modality while the other modality remains largely unused (Wu et al., 2022). In turn, greedy learning negatively affects the model’s generalization ability, which Wu et al. (2022) overcame by proposing an algorithm to balance the conditional learning speeds between modalities during training.
For this study, we investigate the greedy nature of deep neural networks for multi-modal learning from Sentinel-1 SAR and Sentinel-2 MSI data. To that end, we add a Multi-Modal Transfer Module (MMTM) to different levels of the U-Net encoder and decoder. Using squeeze and excitation operations, the MMTM enables intermediate modality fusion for effective multi-modal learning (Joze et al., 2020). The modified dual stream U-Net is then trained on an urban change detection task to investigate the model’s greedy nature using the concept of conditional utilization rate introduced in Wu et al. (2022). The conditional utilization rate for a given modality in a multi-modal setup is the gain on the accuracy when a model has access to the modality in addition to another modality. Finally, we investigate methods to overcome the greediness of multi-modal deep neural networks, including the proposed algorithm in Wu et al. (2022). All experiments are conducted on the SpaceNet 7 dataset, consisting of time series of monthly Planet mosaics and corresponding building footprint annotations for 60 sites covering unique geographies around the world (Van Etten et al., 2021). Sentinel-1 SAR and Sentinel-2 MSI images were downloaded from Google Earth Engine to replace the Planet mosaics (Gorelick et al., 2017). This research is progressing well, and the urban change detection results will be finalized, validated and presented at the Dragon 5 mid-term symposium.

Ground object detection and retrieval from a large-scale remote sensing image are very important to manage smart cities and support sustainable cities. We present a novel problem of visual grounding in remote sensing images. Visual grounding aims to locate the particular objects (in the form of the bounding box) in an image by a natural language expression. The task already exists in the computer vision community. However, existing methods mainly focus on natural images rather than remote sensing images. Compared with natural images, remote sensing images contain large-scale scenes and the geographical spatial information of ground objects (e.g., longitude, latitude). The existing method cannot deal with these challenges. To address the drawback, we design a new method, namely GeoVG. In particular, the proposed method consists of a language encoder, image encoder, and fusion module. The language encoder is used to learn numerical geospatial relations and represent a complex expression as a geospatial relation graph. The image encoder is applied to learn large-scale remote sensing scenes with adaptive region attention. The fusion module is used to fuse the text and image features for visual grounding. We evaluate the proposed method by comparing it to the state-of-the-art methods. Experiments show that our method outperforms the previous methods by a large margin.

In the face of Synthetic Aperture Radar (SAR) image object detection with different distributions of training and test data, traditional supervised learning methods cannot achieve good detection performance. Domain adaptation (DA) method has been shown to have the ability to solve this problem, but existing DA object detection algorithms all use adversarial DA theory for the detection task, which is ineffective in solving object regression localization in the detection task. In this article, to better solve the above problem, an automatic SAR image ship detection method based on feature decomposition crossing different satellites is proposed. The feature extraction layer of backbone network is divided into low level and high level, where domain-invariant feature extractors are designed for the local features extracted from the low level and the global features extracted from the high level, respectively.We argue that the local and global features extracted from source domain and target domain contain domain-specific features (DSF) for adversarial DA and domain-invariant features (DIF) that contribute to object regression localization. Then, we decompose the local features and global features into DSF and DIF via vector decomposition method. For DSF counterpart, we introduce adversarial DA attention for feature alignment. DIF from the local features are fused into the backbone network for high-level global feature extraction. Finally, by using region proposal network and adversarial domain classifier, we can get the accurate bounding box and object class of SAR image objects. Extensive experiments prove that the proposed method outperforms state-of-the-art methods in terms of detection performance.

In the fall of 2021 took place one of the biggest eruptions in the volcanic Canary Islands, on the La Palma island. This paper proposes the use of multispectral Sentinel 2 time series data to monitor pre and post event activity and assess vegetation’s damages. The analysis will be carried out between July 2021 and February 2022. The area of interest is the Cumbre Vieja, an active volcanic ridge on the island. It covers dozens of craters and cones, overlaying the southern half of la Palma island. To study the impact of the volcanic eruption on the vegetation, we will use two vegetation indices, the normalized difference vegetation index (NDVI) and the nonlinear version of this index (kNDVI). The NDVI index is extremely widespread because through it highlitghts changes in vegetation due to both natural disturbances such as wild fires, plant changes and human activities, such as deforestation. Therefore, it can show us how the vegetation was before the eruption and how it looks both during and after the eruption. We expect to find a correlation between the distance from the eruption and changes in the normalized difference vegetation index (NDVI), more precisely we expect to see that NDVI increases with the increase in distance from the eruption. We aim to demonstrate that the nonlinear version of this index (kNDVI) consistently improves accuracy in monitoring key parameters and also that it enables more accurate measures. In the end our goal is to assess these indices against NHI (Normalized Hotspot Indices), a system that performs the automated monitoring of volcanic thermal anomalies.

As an important oceanic physical process, mesoscale eddies play a key role in the processes of ocean mass transport and energy exchange, such as biogeochemical cycles, marine ecosystems, and marine heat balance.

In recent years, these automated ocean eddy identification and tracking algorithms can be divided into two categories: Eulerian- and Lagrangian-based approaches. Based on instantaneous sea surface height (SSH) or velocity field, the major circular structures of mesoscale eddies can be detected, which are called as Eulerian eddies. Previous SSH-based and sea level anomaly-based (SLA-based) methods have shown the best performance due to their ability to avoid extra noise and excess eddy detections. Although SSH-/SLA-based methods worked well at the basin scale, the calculation efficiency decreases distinctly at the global scale, mainly due to the high-order computation complexity on contour iterations. With the improvement of data resolution, existing hardware fails to meet the requirement of long-time-scale global eddy recognition due to the increase of the number of SSH/SLA contours. Thus, based on satellite altimeter data, a highly effective orthogonal parallel algorithm for identifying and tracking global eddies is proposed. Surprisingly, this algorithm is ~100 times faster than the previous SSH-based method on global eddy detection without reducing the accuracy of mesoscale eddy recognition. According to this orthogonal parallel algorithm, the global mesoscale eddy dataset for the past 28 years (1993-2020) was generated, which provides a data foundation for the subsequent study of mesoscale eddies. In terms of the mesoscale eddy dataset, an automatic recognition method of global dipole eddy pairs that consist of two counter-rotating eddies moving together for a period of time within a limited space distance is proposed by using the K–D tree for cutting space. Simultaneously, the transmission modes and characteristics of dipoles are analyzed, including the characteristics of long life, high propagation speed, and entanglement trajectory. In addition, an algorithm named EddyGraph for tracking mesoscale eddy splitting and merging events is proposed and the corresponding data set of eddy trajectories in the Northwest Pacific is available, which would fill the gaps in data sets to support studies on eddy splitting and merging in the Northwest Pacific.

Different from other methods based on instantaneous flow fields, Lagrangian eddies are the cumulative results of the state of the fluid within a given time scale, which can maintain material coherence over the specified time intervals. Firstly, by using the elliptic Lagrangian Coherent Structures, the boundary of a black-hole eddy was extracted based on the data of the geostrophic flow velocity field in the Western Pacific Ocean. Combined with multi-source satellite remote sensing data and in-situ data, the consistency of material transport in the horizontal direction and the coherence of material in the vertical direction of the vortex are analyzed and verified. The results show that the black hole vortex boundary can describe material transport more objectively than the Euler vortex boundary on a longer time scale. Then, the Lagrangian eddies in the western Pacific Ocean are identified and analyzed. By introducing the Niño coefficient, the lag response of the Lagrangian eddy to El Niño is found. Through normalized chlorophyll data, it is observed that Lagrangian eddies can cause chlorophyll aggregation and hole effects. These findings demonstrate the important role of Lagrangian eddies in material transport. Nevertheless, although Lagrangian eddies work well at estimating material transport, the high calculation cost during the integration process has become a bottleneck, especially when the data resolution is improved or the study area is enlarged. Therefore, SLA-based orthogonal parallel detection of global rotationally coherent Lagrangian eddies is built, whose runtime is much faster than that of a previous nonparallel method. Finally, a data set of long-time-scale global Lagrangian eddies is established.

Furthermore, an integrated marine visualization system, named i4Ocean, has been presented. The system is designed and implemented to investigate and study physical marine processes by visualizing and analyzing spatiotemporal marine data. Notably, these actions are realized by providing various GPU-based interaction and visualization techniques for displaying multidimensional data. The system achieves three goals: high visibility, good performance and interactive capabilities. The techniques of z-coordinate calibration and sphere rendering, which restore the most authentic marine environment, provide excellent feedback for oceanographers. The efficient ray sampling technique including a preintegrated transfer function and adaptive sampling methods, increases the rendering efficiency of ocean data. By further introducing a transfer function, users can extract the region of interest in the system and analyze diverse marine phenomena. A data-centric approach was adopted to guide the design of the transfer function by analyzing the scalar field and its properties. The best parameters of the transfer function were obtained to maximize the visibility of important features, which helps to analyze mesoscale eddies of typical ocean phenomena.

Detecting forest insect damage before the visible discoloration (green attacks) using remote sensing data is challenging, but important for damage control. In recent years, the European spruce bark beetle (Ips typographus, L.) has damaged large amounts of forest in Europe. However, it is still debatable how early the infestations can be detected with remote sensing data. Some studies showed a spectral difference between healthy and green-attacked spruce trees at the plot level, while others showed that spectral differences existed before attacks. Therefore, a hypothesis is proposed that no spectral difference can be identified between green-attacked forests compared to healthy forests if the differences do not exist before the attacks. In this study, we tested this hypothesis using Sentinel-2 and WorldView-3 SWIR images on 24 healthy plots and 24 plots with mild, moderate, and severe attacks. In the results, the severely attacked plots did not show significant spectral differences in the Sentinel-2 images until August, and the sensitivity was found in the blue, red, red-edge, and SWIR band. Only the red band showed a significant difference between the healthy and moderately attacked plots in August, and only the blue, red, and SWIR band showed significant differences in September, October, and November. No significant differences were observed in the WorldView-3 images at the plot or individual tree level. We accepted the hypothesis that green attacks do not show spectral differences with the healthy forests when the differences do not exist before the attacks. We concluded that the SWIR bands were sensitive to attacks in the Sentinel-2 images with 10 m resolution, but not in the WorldView-3 images with 3.7 m resolution. Further studies are needed to explore the methodology of using WorldView-3 SWIR images for the early detection of forest infestation.

Accurate and fine multiple tree species supervised classification based on few-shot learning has attracted close attention from researchers, because the sample collection is often hindered in forests. Prototypical networks (P-Nets), as a simple but efficient few-shot learning method, have significant advantages in forest tree species classification. Nevertheless, the overfitting phenomenon caused by the lack of training samples is still prevalent in few-shot classifiers, which brings challenges to training accurate classification models. In this study, we proposed a novel Proto-MaxUp (PM) framework to minimize the issue of overfitting from the perspective of data augmentation and a feature extraction backbone for tree species classification. Taking Gaofeng Forest Farm (GFF) in Nanning City, Guangxi Province, as the study area, nine tree species, cutting site, and road were classified. First, by analyzing the effects of a series of popular data augmentation methods and their combinations in different parts of the P-Net, several effective data augmentation pools were established. Then, the pools aforementioned were combined with PM to obtain the best classification performance. To verify the robustness and validity of the proposed strategy, we applied PM to the other four popular public hyperspectral datasets and achieved excellent results. Finally, this efficient data augmentation method was used in different feature extraction backbones. The results show that the classification accuracy was greatly improved with the optimal backbone (overall accuracy (OA) and Kappa, are 98.08% and 0.9789, respectively), and the difference between training accuracy and test accuracy is less than 2%. It is concluded that the accurate and fine classification for multiple tree species can be realized by the PM data augmentation strategy and backbone proposed in this article.

Pine wilt disease (Bursaphelenchus xylophilus, L.) is one of the major forest diseases in China. The pine wilt nematode infects trees in the Pinus genus, and the infected trees usually die within three months. Since it was discovered in 1982 in China, it has been spreading rapidly and now spread in 19 provinces in China, covering an area of 1.8 million hectares, causing significant damage to the forest and ecological environment. The nematode cannot travel outside the wood independently but is spread by the main insect vector pine sawyer longhorn-beetles (Monochamus spp.,L.) during feeding and oviposition. Therefore, removing the trees infected by the pine wilt nematode from the forest as early as possible is essential to control the spread.

This study aims at developing methods to detect infections using hyperspectral drone images. We inventoried 391 pines in middle east China and recorded them as healthy or early-, middle-, late-stage infected trees. The hyperspectral drone images were obtained with 0.11 m resolution and wavelength from 400 to 1000 nm, covering from red band to near-infrared (NIR). We used the successive projections algorithm (SPA) to select the sensitive bands and the support vector machine (SVM) algorithm to classify trees into different health statuses.

The results showed that when the infection developed into the middle and late stages, the tree crowns showed different signatures from the healthy ones, while during the early stage, the spectral signatures were similar to healthy ones, which decreased the detection accuracy. The classification resulted in high accuracy during the middle and late-stage infection, while separating healthy and early-stage infection was challenging. The spectral signature showed a decreasing ratio between the red and red-edge bands during the infection. We will develop the method using the derivative of the spectral signature to achieve accurate early detection in the future.

Crop classification is one of the most important applications in polarimetric SAR images. Time-series polarimetric SAR images have the characteristics of reflecting the changes of various scattering features of crops in different growth periods. However, since time-series polarimetric SAR needs to combine multiple single polarimetric SAR images, the redundancy between features is multiplied. In this paper, aiming at the problem of feature redundancy, the method of similarity measurement is used to select features from two dimensions of polarimetry and time respectively to reduce feature redundancy. Since the sample size of SAR feature images applied in supervised classification is small, which makes it not suitable for multiple downsampling in CNN, so a suitable classifier based on Transformer is designed. Preliminary experiments on the full polarimetric data verified the effectiveness of the proposed method.

Forest height is an important information for evaluating and analyzing forest resources. Accurate estimation of forest height has great significance for forestry management and ecological research. Multi-band InSAR utilizes the penetration differences of different SAR wavelength, and can obtain the elevations of the underlying topography and the forest canopy surface at the same time. Therefore, it has the ability to extract the forest height. However, in practice, although the long-wavelength InSAR has better penetrating ability than short-wavelength InSAR, its signals are still affected by vegetation scatterers resulting the phase center deviates from the underlying surface. And the extracted elevation of long-wavelength InSAR cannot be directly used as the underlying topography for forest height extraction. On the other hand, the penetration ability of short-wavelength InSAR to vegetation layer cannot be ignored. There is a certain deviation between the phase center and the canopy surface, and the extracted elevation of short-wavelength InSAR cannot be used as a digital surface model (DSM) representing the forest canopy elevation. Therefore, from the perspective of obtaining accurate underlying topography and DSM, the paper proposed a forest height extraction method based on multi-band InSAR data. In this method, an subapeture decomposition approach was used to obtain the underlying topography based on long-wavelength InSAR data. Moreover, based on the gap penetration characteristics of InSAR signal, a DSM penetration bias compensation method for short-wavelength InSAR data was developed based on multi-layer model. Eventually, the forest height was obtained by the difference between the extracted DSM and the underlying topography. To verify the method, the experiment was carried out based on the multi-band airborne InSAR data.

Measuring forest height on a large scale is of importance to forest resource management and biomass estimation. For deformation monitoring applications of radar interferometry, sufficient time series short-baseline polarimetric SAR interferometry (PolInSAR) data have been achieved, and since PolInSAR is limited by interferometric spatial-temporal baseline, using this type of data to estimate forest height will cause a large error. In this paper, a total of five ALOS-2 PALSAR-2 data were obtained in Saihanba forest farm, and the semi-empirical model based on the simplified Random Motion over Ground (RMOG) model and the machine learning algorithm combined with multiple features were used to evaluate the potential for forest height estimation using time series short-baseline PolInSAR data. Experimental results show that: (1) Compared with the semi-empirical model, machine learning algorithms can take full advantage of the multi-feature information of the data and achieve better estimated performance. (2) After combining polarimetric and interferometric characteristics, the phenomenon of overestimation in low regions and signal saturation in high regions can be effectively improved. An R² of 0.44 and RMSE of 3.08m was achieved for inversion result of forest height in pixel size of 90m×90m.

Synthetic Aperture Radar Tomography (TomoSAR) provides an unprecedented opportunity to characterize volumetric environments such as forested areas using 3-D electromagnetic reflectivity maps. Classical 2-D SAR imaging capabilities can be extended to 3-D using acquisitions performed from slightly shifted trajectories, and a coherent synthesis along an additional aperture in elevation. As shown by experiments based on the use or airborne SAR sensors, TomoSAR and its multi-polarization version, PolTomoSAR, is able to characterize various kinds of forests (tropical, temperate, boreal) and may be used to estimates forest height, above ground biomass, underlying ground topography, canopy structure…

However, the application of TomoSAR using spaceborne devices is hindered by the time lag separating successive SAR acquisitions, whose value, on the order of a few days, depend on orbital considerations and on laws of physics. For radars operating at higher carrier frequencies, i.e. at L, S, C, X, Ku bands and above, the correlation time over vegetated environments rarely exceeds hours or minutes, limiting the 3-D analysis through repeat-pass TomoSAR to temporally stable targets, such as those encountered in urban scenarios. A possible solution to this limitation consists in using single-pass interferometers, consisting of two or more SAR sensors measuring, at the same time, the observed scenes from different positions, i.e. in a bistatic configuration. Simultaneous SAR acquisitions permit to solve highly limiting problems related to temporal decorrelation, whereas slight modifications of the relative trajectory between the sensors allows to describe an aperture in elevation and to successfully apply SAR tomographic focusing. Another advantage of this operation mode is that the acquisition of a tomographic stack may be spanned over a large period of time, provided that the structure of the observed medium does not change drastically, i.e. generally months.

This paper illustrates the principles of incoherent bistatic tomography, shows the different processing steps of this technique, which significantly differ from the ones employed to perform repeat-pass TomoSAR. Relevant solutions to operational challenges, linked to the imperfect knowledge of the scene geometry, irregular baseline sampling, and even missing data are presented and validated using airborne data sets using geophysical parameter estimation procedures from the state of the art.

Theoretical aspects will be complemented by an analysis of real data from the ESA campaign TomoSense, where bistatic data were collected at L- and C- Band over the forest site of the Eifel Park, North-West Germany, by flying two airplanes in close formation. Preliminary analyses at L-band were already carried out and gave promising results. A new model-based approach for 3-D reconstruction was developed and implemented; the outcome was then compared to the tomographic profiles produced by standard repeat pass tomography. The structure of the forest was properly recovered in both cases. Quantitative analyses about the visibility of the ground and the forest height were also carried out; the discrepancy of the forest height with respect to the LiDAR map resulted in about 2.3m (1 sigma).