Conference Agenda

The accurate mapping of icebergs embedded in sea ice is important for maritime navigation in the Arctic. Traditionally, large-scale operational monitoring of icebergs in open water has been carried out using C-band SAR sensors. However, recent studies have shown that C-band cannot reliably distinguish between sea ice and icebergs, and therefore C-band is ill-suited for mapping icebergs in areas with high sea ice concentrations. L-band has been shown to solve some of these challenges, as the longer wavelength has a higher penetration into ice and snow, making distinction possible.

In this study, we have tested four different algorithms for detecting icebergs embedded in sea ice. Two machine learning methods, a logistic regression classifier, and a support vector machine classifier, were implemented, with added textural features to aid the classification. These were compared against a traditional Constant False Alarm Rate (CFAR) method serving as a benchmark. In addition, a novel multiscale CFAR algorithm was introduced to address the issues arising from the large size differences of the icebergs in our study area.

All four algorithms were tuned using a single image and then applied to 4 new images covering two distinct years with different ice conditions. Validation was carried out using independently obtained iceberg positions from optical satellite images. For most images, the machine learning methods slightly outperformed the classical CFAR method, with the main drawback being the computation times for calculating the textural features. The multiscale CFAR method was able to outperform the machine learning methods without the need for textural features and model training. The machine learning methods, however, show some advantages over the multiscale CFAR method, e.g., their application resulted in a higher number of true positives.

The study highlights the advantages of the novel multiscale CFAR algorithm but also underscores the utility of using L-band SAR for detecting icebergs in areas with a high sea ice concentration, potentially enhancing navigational safety in Arctic waters. The results of the machine learning methods, although outperformed by the multiscale CFAR, do show some promise in terms of high recall, indicating the need for further studies.

In conclusion, the study contributes valuable insights into the ongoing efforts aimed at improving operational iceberg detection in the Arctic, leading to safer maritime navigation.

The Arctic sea ice exerts significant impacts on global climate change, navigation, shipping, and natural resource development, while also influencing the detection of other marine phenomena. Therefore, sea ice monitoring holds crucial research significance and practical value. Sea ice identification and classification are fundamental components of sea ice monitoring. SWIM, a novel small incidence angle sensor mounted on CFOSAT, offers high data coverage but lacks imaging capability and has relatively low spatial resolution. Hence, this study employs image reconstruction techniques to obtain high-quality sea ice images. Present image reconstruction techniques primarily rely on medium incidence angle microwave scatterometers, which enhance spatial resolution by extracting information from overlapping regions between multiple independent measurements in the same area. This study proposes image reconstruction under small incidence angles based on SWIM data. Comparison of reconstruction results over one, three, seven, and ten days reveals that seven days of data achieve complete coverage of the entire Arctic region. For image reconstruction, SWIM data within the incidence angle range of 6° to 10° are selected. Through systematic investigation and comparison, it is found that at 6°, the reconstructed image exhibits numerous false bright spots; at 8°, the number of bright spots significantly decreases, and finally, at 10° incidence angle, the reconstructed image displays the fewest false bright spots, indicating that data with a 10° incidence angle are more suitable for image reconstruction under small incidence angles. Considering the issue of low footprint coverage under a single incidence angle, this study fuses data within the 6-10° incidence angle range after small incidence angle correction for image reconstruction. Based on the reconstructed images, sea ice identification and classification are conducted, and sea ice edges are extracted. Comparative validation against previous studies on SWIM sea ice edge extraction, NSIDC products, AARI products, SSM/I Tb sea ice edge images, ASCAT images, Sentinel-2 images, etc., demonstrates the high accuracy of the reconstructed sea ice edges. In sea ice classification, this study utilizes data from three individual incidence angles, with results indicating that classification accuracy is highest at a 10° incidence angle, achieving an overall classification accuracy of 89.1%. This study fills the gap in research on image reconstruction techniques under small incidence angles, promoting the development of sea ice remote sensing detection technology and applications. It holds significant theoretical importance and practical value for polar sea ice monitoring, ice situation assessment, and forecasting.

Melting of sea ice in the Arctic has accelerated due to global warming. The Fram Strait (FS) serves as a crucial pathway for sea ice export from the Arctic to the North Atlantic Ocean. Monitoring sea ice drift (SID) in FS provides insight into how Arctic sea ice responds to the climate change. The SID has been retrieved from Sentinel-1 SAR, AVHRR, MODIS and AMSR-E, and further exploration is needed for the retrieval of SID using optical imagery. In this paper, we retrieve SID in the FS using the Coastal Zone Imager (CZI) on Chinese HaiYang-1D (HY-1D) satellite. Multi-template matching technique is employed to calculate cross-correlation, and subpixel estimation is used to locate displacement vectors from the cross-correlation matrix. The dataset covering March to May 2021 was divided into hourly and daily intervals for analysis, and validation was performed using Copernicus Marine Environment Monitoring Service (CMEMS) SAR-based product and IABP buoy. A comparison with the CMEMS SID product revealed a high correlation with the daily interval dataset; however, due to the spatial and temporal variability of sea ice motion, differences are observed with the hourly interval dataset. Additionally, validation with IABP buoy yielded a velocity bias of -0.005 m/s and RMSE of 0.031 m/s for the daily interval dataset, along with a flow direction bias of 0.002 rad and RMSE of 0.009 rad respectively. For the hourly interval dataset, the velocity bias was negligible (0 m/s) with a RMSE of 0.036 m/s, while the flow direction bias was 0.003 rad with a RMSE of 0.010 rad. In addition, during the validation with buoys, we found that the accuracy of retrieving the SID flow direction is distinctly interrelated with the sea ice displacement.

By utilizing Polarimetric Synthetic Aperture Radar (PolSAR) data, multiple polarization features can be extracted, thereby offering comprehensive electromagnetic scattering characteristic information for sea ice. However, the classification capabilities of multi-frequency polarimetric features for sea ice during the melting season have not been thoroughly discussed. Furthermore, although various classifiers have been applied to sea ice classification, the impact of different combinations of classifiers and feature sets on sea ice classification results remains unknown. This study addresses the intricacies of sea ice characteristics during the Bohai Sea melting period, conducting the inaugural evaluation of sea ice classification capabilities utilizing L, S, and C-band full-polarization PolSAR data. The study categorizes Bohai Sea sea ice during the melting period into five classes: Open Water (OW), Grey Ice (Gi), Melting Grey Ice (GiW), Grey-White Ice (Gw), and Melting Grey-White Ice (GwW). Initially, 51 polarimetric features are extracted from L, S, and C-band PolSAR data using diverse polarization decomposition methods. Subsequently, the Euclidean distance (ED) of these polarimetric features among different sea ice type combinations is calculated, assessing their separability and analyzing discrepancies. Finally, the influence of various feature combinations on classifier performance is examined, identifying the optimal classifier-feature set combination. Experimental findings reveal that total power parameters (Shannon entropy (SE, SE_I), scattering matrix total power (Span)), volume scattering parameters (Freeman decomposition volume scattering corresponding power (P_V-Freeman)), and scattering mechanism parameters (eigenvalues (λ_i)) exhibit robust classification capabilities across the three bands. Comparative analysis of polarimetric features shows that the L-band excels in classification for Gi-GiW, Gi-Gw, GiW-Gw, and GiW-GwW combinations. The S-band demonstrates superior classification performance in the Gi-GwW combination, while the C-band showcases the highest classification ability in OW-Gi, OW-GiW, OW-Gw, OW-GwW, and Gw-GwW combinations. Specifically, within single-band polarimetric feature sets, OW and Gi attain the highest classification accuracy in the C-band, GiW in the S-band, and Gw and GwW in the L-band. Overall, the classification accuracy of C-band polarimetric features surpasses that of L and S bands. Employing a multidimensional polarimetric feature set yields a classification accuracy of 94.55%, representing a notable enhancement of 9% to 22% compared to single-band classification. Opting for the Random Forest (RF) classifier achieves the highest classification accuracy, with the optimal multi-dimensional feature composition comprising: L-band: SE, SE_I, α ̅, Span; S-band: SE_I, SE, Span, PV-Freeman, λ₁, λ₂; C-band: SE, SE_I, Span, λ₃, P_V-Freeman, reaching 95.84%. The outcomes of this study introduce a robust approach for future sea ice monitoring during the melting season.

Launched on December 16^th, 2022, the innovative SWOT mission provides observations of inland water bodies at unprecedented resolution and accuracy. The surface water areas estimated shall have a relative error smaller than 15% for water bodies exceeding 250*250 m², and the height accuracy shall be 10 cm or better for water bodies exceeding 1 km² and 25 cm or better for water bodies between 250*250 m² and 1 km². The preliminary results obtained over the Rhine Tier 1 CalVal site are within the requirements with margin, it was interesting to test it over lakes known for their great inter and intra annual dynamic variations as well as lakes setting in a different environment.

An assessment of SWOT L2 Lake HR products has been carried out over the small lakes of the Anhui province Caizi, Wuchang , Shengjin and Baiding with ranging from 40 to 200 km², as well as the sub lakes of Dongting (Caisang, Daxiaoxi and Chunfeng) and Poyang (Zhonghuchi, Banghu, Shahu, Dahuchi, Dachahu, Meixihu). SWOT data are acquired since the mid July 2023; a site is overpassed at least every 21 days, with for some ones, up to 4 observations by cycle.

A preliminary work consisted in setting up a database containing water surface dynamics based on 6 years of Sentinel-2 images, water elevation from historical data and also altimetric data derived from ICESat-2.

The analysis of SWOT products was carried out as follows:

- Evaluation of the accuracy of the PIXC classes (water only, land and water, dark water, etc).

- Exploitation of Bright Land and Dark Water flags.

- Comparison of the in situ water surface elevations (wse) and the ones provided by the L2 SWOT lake products.

- Comparison of water surface areas (wsa) provided by the SWOT product and the ones derived either from quasi-synchronous HR Sentinel-2 data.

Obtained preliminary analyses show wse and wsa errors lower than the SWOT requirements, highlighting the high accuracy of the SWOT products.

In the past 20 years, under the influence of global climate change and human activities, the relationship between rivers and lakes in the middle and lower reaches of the Yangtze River has continued to be adjusted, and the hydrological rhythm of Poyang Lake has changed dramatically. Climate change has led to extreme hydrological events in Poyang Lake. In 2020, over-standard flooding occurred in the Yangtze River, and in 2022, the Yangtze River Basin experienced a historically rare event of " returning to dryness during flood season". During November 2022, Poyang Lake suffered from a severe drought disaster, and the water level at Xingzi Station receded to 6.46 meters, which set a new record low water level. Since 2023, the overall water level of Poyang Lake has continued to be low.

In order to explore the impacts of these extreme hydrological events on the hydrological rhythm of Poyang Lake, we used multi-source remote sensing satellite data (Sentinel-1, Sentinel-2, ICESat-2, etc.) to carry out high spatial-temporal resolution and all-weather monitoring of key hydrological elements (water level, water area, water storage, etc.) , and verified the results with the hydrological station data. 1) We found that the water level and water area showed strong correlation in recent years, especially at Xingzi station (R²=0.88). 2) The optimal RMSE between ICESat-2 data and hydrological station data after error correction is 0.625m, furthermore, we realized the detection of water level change in seasonal lakes of Poyang Lake by ICESat-2. Under the background of the lack of hydrological stations and continuous monitoring data, we characterized the change pattern of water level and area of seasonal lakes in recent years, which provided data support for the changes of food and habitat environment of migratory birds. 3) For purpose of assessing the drought disaster in Poyang Lake more accurately, we carried out the research on the precise classification of land cover. Our research results can provide decision support for the relevant management departments for disaster early warning and assessment of Poyang Lake.

Plant communities play an important role in wetland elements and are vulnerable to human activities and climate change. Wetland plant community classification and mapping provide scientific important data support for wetland ecological monitoring and evaluation. This study aims to develop a classification scheme suitable for the wetland plant communities in the Poyang Lake wetland. Taking Poyang Lake National Nature Reserve as the research area and on the basis of the monthly Sentinel-1 and Sentinel-2 time-series data in 2019, this study extracts five types of image feature parameters, including water and vegetation index group, red edge index group, texture feature group, spectral feature group, and polarization radar backscatter group, with a total of 240 feature indexes, and uses Random Forest (RF), Support Vector Machine (SVM), and Deep Neural Network (DNN) algorithms for classification to explore a set of optimal feature combinations and a suitable classification scheme for wetland vegetation mapping in Poyang Lake. Results show the following: (1) Compared with radar data, the extraction accuracy of optical data is remarkably better than that of radar data in wetland plant community classification and mapping. Radar data can be used as a supplement to optical data when optical data are insufficient. (2) Screening the importance of each image feature of Sentinel-2 helps improve the classification accuracy. The preferred time periods are mainly distributed in January, May, August, September, October, and December. (3) Five groups of unitary image features are selected to classify separately, and the classification accuracy is as follows: red edge index group > water and vegetation index group > spectral feature group > radar polarization data group > texture feature group. (4) Comparing the combined image feature groups with the unitary image feature groups reveals that the combined image feature group is not necessarily helpful to improve the classification accuracy. The classification accuracy is as follows: red edge index group > water vegetation index group > combined image feature group. Among them, the overall accuracy of the classification scheme using the red edge index group and random forest method is 0.81, and the Kappa coefficient is 0.76. (4) By comparing the three classification algorithms, the classification accuracy is ranked as follows: DNN > RF > SVM. The overall accuracy of the deep learning method does not greatly improve, that is, only 2% higher than the RF algorithm. Thus, the DNN and machine learning method (RF) can be used as optimization algorithms. In conclusion, A classification scheme for wetland plant communities in the Poyang Lake wetland was proposed in this study using multi time-series Sentinel-2 and Sentinel-1 data. The optimal acquisition time periods of satellite data are in January, April, August, September, October, and December. The optimal image feature group can be red edge index group or water and vegetation index group for feature selection. The classification algorithm can select deep learning or RF algorithm to classify wetland plant communities according to the requirements. This classification scheme can effectively improve the accuracy of wetland vegetation mapping in the Poyang Lake and provide scientific and technical solutions for decision-making departments.

Soil moisture (SM) and evapotranspiration (ET) are critical components of the water-energy-carbon cycle, and their accurate representation in land surface models (LSMs) is essential. Integration of satellite-based observations of SM and leaf area index (LAI) provide valuable information and could reduce model prediction uncertainty, improving the characterization of terrestrial states and fluxes in LSMs. In this work, the iterative Ensemble Smoother (IEnS) is used to estimate sensitive vegetation and soil model parameters and states by assimilating the Soil Moisture Active Passive (SMAP) soil moisture, Moderate Resolution Imaging Spectroradiometer (MODIS) LAI data and evapotranspiration data measured at ICOS sites into a land surface model (Community Land Model 5.0). The proposed framework is implemented over the EUROCORDEX domain with a computational grid size of 3 km and for the assimilation period of 2018-2020. The simulations with updated parameters were assessed using in-situ observations, including SM measurements from the Terrestrial Environmental Observatories (TERENO) network and ET measurements from the Integrated Carbon Observation System (ICOS). Furthermore, we investigated the impact of parameter updates on gross primary production (GPP) and net ecosystem exchange (NEE). Results show that the assimilation of remotely sensed soil moisture improves the simulation of soil moisture over the European domain, while evapotranspiration simulation is hardly improved. ET measured at ICOS sites is used to update plant functional type (PFT) specific vegetation parameters, for example those which control stomata resistance and plant hydraulic water uptake, and it is shown that this both improves SM and ET modelling. Results illustrate the importance of assimilation of vegetation related measurements and estimation of vegetation parameters to improve the characterization of soil moisture and ET.