Anthony Paul Doulgeris

We present a robust, incidence angle (IA) and wind speed aware model for separating sea ice from open water and providing high resolution sea ice concentration (SIC) estimates from L-band synthetic aperture radar (SAR) imagery, with potential climate study and operational benefits. By treating open water as a wind speed-dependent class in an established IA-aware Bayesian maximum likelihood ice type classifier, we account for the wind-driven open water backscatter variability by using external wind information. This method circumvents the need for the extensive training datasets or accurate geophysical model functions (GMFs) typically required by established C-band ice/water separation algorithms. To demonstrate this approach, we utilize wide-swath dual-polarized (HH/HV) L-band SAR data from the ALOS-2 mission. The proposed method provides high accuracy during the challenging melt period, achieving high Matthews Correlation Coefficient (MCC) scores (MCC>0.800), and consistently outperforms passive microwave radiometer (PMW) products during this period. During winter conditions, weak sea ice backscatter and system noise limitations hindered reliable ice/water separation (MCC=0.506). We compare our proposed method to a baseline model without incorporated external wind information, and find that the wind-informed classifier achieves consistently higher classification accuracies. Our results demonstrate that integrating external wind data allows for robust, high resolution ice/water separation in L-band SAR imagery during melting conditions based purely on backscatter intensity, removing the need for computationally heavy textural features.

Accurate and high-resolution sea ice concentration (SIC) products are needed for operational forecasting and climate monitoring. Passive microwave sensors, like the Advanced Microwave Scanning Radiometer 2 (AMSR2), provide daily SIC observations at 5 km resolution. However, this coarse spatial resolution constrains their capabilities. Whereas, Sentinel-1 Synthetic Aperture Radar (SAR) offers high-resolution ice-water discrimination but is prone to classification errors in wind-roughened open water and land-fast ice. This study aims to obtain the strengths of both sensors to produce a robust km-scale SIC for winter Arctic conditions by applying a SAR–AMSR2 data fusion framework using a U-Net model trained on high-resolution pixel-level labels. The U-Net uses SAR features (HH, HV, incidence angle) and AMSR2 inputs, with thirteen feature combinations across two fusion strategies being evaluated: direct data-level fusion of AMSR2 brightness temperatures or preprocessed data-level fusion of AMSR2 SIC. The findings show that fusing AMSR2 SIC with SAR improves classification with mean intersection over union scores of 97.93% and 92.47% compared with SAR- and Landsat-derived validation data, with significant improvements in previously identified troublesome land-fast ice and open water regions. The preservation of spatial resolution in the fused SIC is also evaluated by using power spectral variance analysis, which confirms that the fusion does not degrade the effective spatial resolution of the final output. Overall, the approach presented in this study supports accurate, high-resolution winter SIC retrieval suitable for future operational forecasting and data assimilation, like the Barents-2.5 km model, by preserving fine-scale features and reducing misclassification.

This study proposes a multiscale extension of the well-known constant false alarm rate (CFAR) outlier detection method for the detection of icebergs in L-band synthetic aperture radar (SAR) images. The proposed method offers the possibility to accurately detect both single icebergs or clusters of icebergs embedded in sea ice. The method was tested on 8 L-band dual-pol SAR images acquired at various incidence angles and compared against a conventional single-scale CFAR approach. Ground-truth iceberg locations were derived from optical satellite images to ensure accurate validation of the results. The results reveal that the multiscale method improves the detection performance for icebergs of all sizes with an increase in recall of 0.05 and precision of 0.04 on average, leading to an increase in F1 score of 0.05 on average. In addition, the suggested multiscale method improves the estimation of the total iceberg area compared to a single-scale method, with a relative error of the estimated total area of 0.53 on average, compared to 1.05 for the multiscale approach. The results demonstrate the feasibility of using a multiscale outlier detection method on SAR images for the operational detection of icebergs in sea ice.

We present a methodology to derive sea ice type classification maps in the Atlantic sector of the Arctic Ocean for the autumn through spring seasons, from 1991 to present. We use datasets from four C-band Synthetic Aperture Radar (SAR) sensors: Sentinel-1, RADARSAT-2, Envisat ASAR, and ERS-1/2. We demonstrate a classification workflow to specifically accommodate different characteristics of these datasets to produce satisfactory and consistent classification results. We use a Gaussian sea ice type classifier correcting for per-class incidence angle effects on SAR images. Input features include backscatter intensities corrected for thermal noise mainly using Kalman filtering, and Gray-level Co-occurrence Matrix (GLCM) based image textures. Class labels include open water, young ice, level and deformed first-year ice, and multi-year ice. A robust Markov Random Field (MRF) based contextual smoothing process is applied to produce final classification maps. Assessments of classification accuracies and consistency of class fractions between datasets for the training scenes, as well as for a one-year classification time series in a test area and one-month classification map mosaics over the Atlantic Arctic, suggest that our workflow is capable of producing temporally consistent large-scale ice type maps at sub-km spatial resolution. This study provides the theoretical foundation for a three-plus decade C-band SAR-based sea ice type classification product. Such a product will contribute to the assessment of seasonal and inter-annual sea ice variations and climate trends, which is critical for understanding physical and biogeochemical processes acting across the Arctic's ocean-ice-atmosphere interface.

Accurate and high-resolution Sea Ice Concentration (SIC) mapping is essential for polar navigation, environmental monitoring, and assimilation into forecast models. Traditional passive microwave sensors, like AMSR2, provide reliable SIC estimates but are limited by coarse (5 km) resolution, particularly near coastlines and regions with mixed ice and open water, where finer spatial detail is critical. Synthetic Aperture Radar (SAR) imagery offers a high-resolution alternative. This study applies a U-Net convolutional neural network to Sentinel-1 SAR data, utilising pixel-wise ice-water labels to enhance SIC mapping. To address SAR noise challenges, we incorporate multilooking, adaptive noise correction, and overlapping patches at inference to improve SIC accuracy while preserving fine-scale features. We trained the U-Net across multilooking levels to balance resolution, noise reduction, and computational efficiency, allowing the model to handle noise artefacts effectively. Our results identify an optimal 7 × 7 multilooking level, achieving 280 m ice-water labels and a 2.5 km SIC field when an additional 9 × 9 SIC window is applied. This configuration enhances traditional SIC products by improving the representation of the ice edge, leads, and nearcoastal features, which are critical for operational applications. SAR-derived SIC addresses the limitations of passive microwave products by providing superior spatial detail and ice edge resolution. Incorporating additional information from AMSR2 or wind features could strengthen SIC robustness and minimise the misclassification of open water, which is present in the results. These advancements would establish SAR-based SIC as a valuable tool for operational sea ice monitoring and integration into high-resolution models.

We present a feasibility assessment of using several publicly available C-band wide-swath SAR datasets to derive sea ice type maps in the Atlantic sector of the Arctic Ocean from 1991 to present. This region is characterized by highly variable and dynamic sea ice conditions, and temporally consistent, large-scale monitoring of sea ice parameters is only possible through satellite remote sensing. We use data from C-band sensors including Sentinel-1, RADARSAT-2, Envisat ASAR and ERS-1/2, which have similar central frequencies and spatial resolution, to cover the study period. We evaluate comparative image classification performances and classification consistency using these datasets with common training datasets in geographically and temporally overlapping scenes and a sea ice classifier correcting for per-class incidence angle (IA) effects. Through this evaluation, we demonstrate the differences in these datasets affecting sea ice classification and the feasibility of using legacy sensors including Envisat ASAR and ERS-1/2 to extend the time series of sea ice type maps back to 1991 in our study area. This study provides theoretical support for the establishment of a multi-decadal SAR-based sea ice type product, which will contribute to the assessment of seasonal and inter-annual sea ice variations, especially the variability in new ice formation, which strongly influences physical and biogeochemical processes across the ocean-ice-atmosphere interface. This study is part of the collaborative project INTERAAC (air-snow-ice-ocean INTERactions transforming Atlantic Arctic Climate) between Norway and China, which aims at generating a reconciled multi-mission Climate Data Record (CDR) for Atlantic Arctic sea ice.

We estimate sea-ice type specific incidence angle (IA) dependencies for dual polarized (HH/HV) L and C-band synthetic aperture radar (SAR) for the winter, melt onset and advanced melt seasons for level and deformed ice, using time-series of Advanced Land Observing Satellite-2 (ALOS-2) and Sentinel-1 imagery off the north-east coast of Greenland. The IA dependencies are used to radiometrically correct the L and C-band backscatter time-series, which enables analysis of their seasonal evolution. From this, we observe that the L-band backscatter intensity increases for both ice types at the transition from winter to melt onset. We use these results to estimate ice type separability and to train an IA aware Bayesian classifier at both frequencies. These results show that while both frequencies are capable of distinguishing level and deformed ice during the winter, only L-band SAR can reliably make this separation during the melt onset season. During the advanced melt season, the overall classification accuracies are similarly low for L and C-band. This study demonstrates the potential of L-band SAR for sea-ice mapping, which is highly relevant in the light of several upcoming L-band SAR missions.