Conference Agenda

Traditional climate risk and impact assessments typically consider a single extreme event, a fact that leads to the underestimation of risks, as such events are often interdependent. The principal aim of this study is to evaluate the current state of the climate in Beijing, China, and Athens, Greece in terms of droughts and heatwaves, focusing on their compound effects (CDHW) and examining their association with urban form and fabric factors. The term compound events describe the combined effect of multiple climate factors (processes, variables, phenomena including feedback mechanisms) or climate hazards. In urban areas, these compound events can lead to other challenges, such as increased energy demand for cooling, higher air pollution levels, and impacts on critical infrastructure which can be associated with urban morphology. The determination of the CDHW climatology is carried out through the joint use of an Excess Heat Factor (EHF) and a Standardized Precipitation Index (SPI), according to the general definition of CDHW events (heat waves occurring during the period of drought events), using the high-resolution state-of-the-art ERA5-Land reanalysis product along with ground-based climate data, while Earth Observation (EO) imagery is used to extract land cover information from visible and near-infrared sensors. The study addresses the challenges of CDHW in cities and a range of strategies is proposed that include climate-resilient infrastructure, nature-based solutions, and heat warning systems.

While more and more people migrate to cities, uncontrolled urban growth poses pressing threats such as poverty and environmental degradation. Although sustainable urban planning can mitigate these threats, the lack of timely information on the sprawl of settlements hampers ongoing sustainability efforts. Multi-modal deep learning offers new opportunities for timely and accurate urban mapping and change detection by exploiting the complementary information acquired by Synthetic Aperture Radar (SAR) and optical sensors. In particular, the Copernicus Program's Sentinel-1 SAR and Sentinel-2 MultiSpectral Instrument (MSI) missions play a crucial role in multi-modal remote sensing research. For example, our previous work demonstrated that the complementary information in Sentinel-1 SAR and Sentinel-2 MSI data can be utilized to improve the transferability of deep learning models for urban extraction at a global scale (Hafner et al., 2022). However, the optical modality may not always be available due to cloud cover or other atmospheric conditions, which is particularly relevant for multi-temporal urban mapping and change detection. Although a limited number of studies have addressed this so-called missing modality problem (e.g., Zheng et al., 2021, Saha et al., 2022, and Li et al., 2022), multi-modal methods that are robust to a missing modality are still under-researched in remote sensing. Here, we propose a novel multi-temporal urban mapping approach that uses multi-modal satellite data from the Sentinel-1 SAR and Sentinel-2 MSI missions. In particular, our approach focuses on the problem of a partly missing optical modality due to clouds. The proposed model utilizes two networks to extract features from each modality separately. In addition, a reconstruction network is utilized to approximate the optical features based on the SAR data in case of a missing optical modality. Our experiments on a multi-temporal urban mapping dataset with Sentinel-1 SAR and Sentinel-2 MSI data demonstrate that the proposed method outperforms a multi-modal approach that uses zero values as a replacement for missing optical data, as well as a uni-modal SAR-based approach. Therefore, the proposed method effectively exploits multi-modal data, if available, but it also retains its effectiveness when the optical modality is missing.

Self-supervised learning methods are gaining popularity in remote sensing community due to their ability to utilize unlabeled data for representation learning. These representations can then be adapted to downstream tasks through pre-training and fine-tuning. Masked Autoencoder (MAE) is a concise self-supervised learning method that learns better semantic representations by masking most of the content in the input image. However, MAE was originally designed for natural images and may not be the best choice for remote sensing images. We propose a masking method to enhance correlation feature extraction capability. Our proposed model surpasses state-of-the-art contrastive learning and MAE-based models on land-cover classification tasks and reduces input data volume, achieving a more efficient model. Additional experiments demonstrate that the proposed model has good generalization performance and maintains good representation learning capabilities on small-scale data.

Artificial Intelligence (AI) is currently studied mainly for optical imagery, i.e. photography. Earth Observation (EO) images are basically different and much more complex. AI for EO requires specific methods for the full information extraction from spatial, temporal or spectral information at global scale. This involves new paradigms to analyze jointly multimodal sensor records as the EO multi-sensor data optical, IR or microwaves. EO records data of high complexity, physically-based, dynamic, non-linear coupled Earth System. We need to develop new AI paradigms with integrated physical principles into the learning mechanism. These are well beyond and do not emerge form the present cats and dogs recognition techniques. Thus, there is a huge motivation in developing AI for EO methods and exploiting the results.

Eddy interactions, including typical splitting and merging processes, are a popular research focus in oceanography. Automatic splitting and merging identification algorithms are crucial for global eddy interaction research. This study proposes an algorithm for identifying and tracking global mesoscale eddy splitting and merging events based on sea level anomaly (SLA) data. For identification, we present a multilevel eddy detection method that introduces eddygroups and eddytrees to describe the complicated spatial and topological relationships between different levels of closed SLA contours. For tracking, we define an eddy segment, eddy branch and eddy directed acyclic graph (eddy-DAG) to describe the complex topological trajectory of eddies that include at least one splitting or merging event. Only eddies contained within a common eddygroup and with the same polarity can be tracked as sources for merging events or sinks during splitting events. The Global Eddy Graph dataset (DOI: 10.12237/casearth.63369940819aec34df2674d8) extracted 1,905,742 splitting events as well as 1,790,266 merging events from CMEMS’s SLA data (1993-2020). Based on the typical events extracted from the Global Eddy Graph, the normalized results of different remotely sensed sea surface parameters (SSTA, SSSA) or in situ data (drifters) verify the reliability of the dataset and the effect of the interaction between eddies on marine material distribution.

As an important place for shipbuilding enterprises to manufacture and repair ships, docks and berths are the most critical components of shipbuilding enterprises.In the shipyard scene, the dock and berth are closely related to the production status of the shipyard. They are the core land types in the shipyard production status monitoring. Therefore, the production status of the shipyard can be inferred by monitoring the dock and berth in the satellite remote sensing image.In this paper, based on the characteristics that shipyards with different production states differ greatly in remote sensing images, five deep learning networks ( GoogLeNet, integrated network, Xception, VGG and Alexnet ) are used to train and predict the dock data set, and the accuracy and effect of the evaluation model are compared. Then, combined with the shipyard vector data, the production state activity of the shipyard 3km along the coastline is counted. The experiment adopts cross-time series statistics, and selects the areas with different production state activity across time series as the research area ( the research area chooses to avoid factories and many housing construction areas ). Finally, the Sentinel-2A image data of the selected study area in the cross-temporal period was obtained, and the water body was extracted by MNDWI. The water color index (FUI), turbid water index (TWI), cyanobacteria and macrophytes Index (CMI), river pollution index (RPI) were calculated to evaluate the water pollution situation, and the correlation analysis between the activity of the shipyard and the water pollution situation was established.

Landslides are one of the most serious geological disasters in the world, which seriously damage people's property and safety. In this paper, an object-oriented segmentation method is proposed, which combines spectral, terrain and texture features. The Lengqu basin, a tributary of the Nujiang River on the south side of the Tanggula Mountains in China, and parts of the Hunza basin in Pakistan were selected as the study areas. Landslides in the study area were identified using 12.5 m elevation data and Sentinel-2 data. The identification results were validated against images on Google Earth and collected landslide data. The results show that the object-oriented method can extract the landslide boundary accurately. The research results have great scientific significance for disaster prevention and mitigation, line planning and site selection and follow-up maintenance of the Sichuan-Tibet transportation corridor and the Karakorum line.

Remote sensing techniques are widely used for identification of ancient landslides and monitoring their activity In the present study, we used the Hunza valley basin in Pakistan as the study area, and enhanced the DEM (Digital elevation model) based on RRIM (Red relief image map) to identify the ancient landslides the SBAS InSAR (Small baseline subset synthetic aperture radar) technique was also used to monitor the surface deformation rate in the study area from 2004 to 2022 and then the histograms of the radar line of sight deformation rate results were used to categorize the deformation rate results In this research, a total of 157 ancient landslides with activity characteristics were identified It is found that the RRIM method supplemented with InSAR technology can effectively monitor the ancient landslides and avoid the risk by monitoring the hidden ancient landslides in a long time series.

The Karakoram region is located on the tectonic belt and is also a high-risk area for geological disasters. Due to the complex terrain, high mountains and deep valleys, geological disasters such as landslides are prone to occur, and traditional monitoring is extremely difficult to carry out, which hinders the understanding of landslides in the region and leads to a lack of disaster prevention and reduction measures for local landslide disasters. This study is based on the 2021 Sentinel-1A data along the Karakoram Highway, and starts from the identification results of Stacking InSAR technology, focusing on analyzing typical landslides along the Karakoram Highway. Utilizing Small Baseline Subset Synthetic Aperture Interferometric Radar (SBAS InSAR) technology to monitor the displacement characteristics of landslides, and analyzing the causes of landslides in conjunction with the environment in which they occur. The research results are as follows: (1) Based on Stacking InSAR technology, 7 potential landslides along the Karakoram Highway were obtained, all of which are in an unstable state. (2) In 2021, landslides occurred frequently along the Karakoram Highway, and the displacement data of the landslide line of sight showed significant deformation of the Mostag landslide, with a maximum deformation rate of 94 mm/a. The research results are of great significance to the prevention and control of geological disasters along the Karakoram Highway and to serving the national "the Belt and Road" strategy.

A good-quality landslide inventory map is necessary for assessing landslide hazard. However, it remains difficult and time-consuming to produce and update landslide inventories in most regions of the world, especially in mountainous areas with high extension and poor accessibility. Moreover, the inventoried landslides are usually the most morphologically visible on the landscape, while other typologies of large dimensions and more diffuse boundaries are often overlooked. Therefore, new technologies such as satellite remote sensing or advanced landscape analysis are gaining prominence to optimise landslide mapping at regional scale, in terms of time-consuming and cost-effectiveness. In this study, we performed a combination of two well-implemented techniques to improve landslide detection in a mountainous area. These techniques are Differential Interferometric Synthetic Aperture Radar (DInSAR) and Landscape Analysis through the double normalised channel steepness (k_sn) geomorphic index. The southwestern sector of Sierra Nevada mountain range (Granada, Southern Spain) was selected as the case study.

We derived DInSAR mean displacement or velocity maps from Sentinel-1 images through the P-SBAS automated and un-supervised processing chain, that is implemented on the European Space Agency (ESA)’s Geohazard Exploitation Platform (GEP) (https://geohazards-tep.eu/#!). Ascending and descending orbit data was obtained with spanning times from September 2016 to March 2020 and December 2014 to March 2020, respectively, with temporal sampling up to 6 days. The k_sn index was computed through the open Python library ‘landspy’ (https://github.com/geolovic/landspy). The only needed input was a 10 m resolution Digital Elevation Model to extract the drainage network and the k_sn index from rivers.

We identified the unstable areas from the DInSAR ground displacement maps and the k_sn anomalous values from the k_sn map to associate them with large landslides. To delimit the landslides’ boundaries as accurately as possible, it was essential an exhaustive examination of morphologies in the field, as well as the examination of products derived from high-resolution Digital Elevation Models (e.g. hillshade, slope, aspect, rugosity). This work conducted us to provide an updated inventory of 28 landslides, what implies the 33.5% of the analysed area. Most of the identified landslides are large Deep-Seated Gravitational Slope Deformations (DGSDs), that have not been discovered in the Sierra Nevada until this study. This new inventory has relevant implications as landslides are larger and more abundant than previously considered. Our work also emerges the potential of integrating data from DInSAR techniques and Landscape Analysis to detect large landslides and provide updated inventories in mountainous areas. Moreover, we proved that some limitations of both techniques could be well-compensated.

Monitoring mining subsidence dynamically offers valuable opportunities for exploring and examining the directional changes in surface displacement resulting from underground resource extraction. These changes can be significantly influenced by both natural geological environmental factors and human activities. LuTan-1(LT-1) mission is the first L-band bistatic spaceborne SAR mission for civil application in China which provides continuous DInSAR ground deformation results. Although orbital determination accuracy of LT-1 is 5 cm, we conducted a linear fitting and removed the orbital-induced phase ramp by means of Kriging’s interpolation method in this work. The subsidence bowl results derived from LT-1 show good agreement with the results derived from Sentinel-1 between March and April 2022. Furthermore, due to the scarcity of GNSS points and the irregular mining deformation, it is difficult to obtain high precision 3D deformation through GNSS and InSAR. Therefore, we projected continuous 3D GNSS to LOS direction to validate the DInSAR results derived from LT-1 and Sentinel, respectively. Finally, we observed the dynamic process associated to mining activities in this area by using four DInSAR results from different dates. InSAR results revealed obvious directional changes of the spatial location of ground surface displacements, with maximum horizontal displacement of the subsidence bowl of about 1.26 km during the observation time lag of approximate one year. This approach opens the door to the dynamic analysis of mining subsidence by DInSAR method.