Hongyan Xi

The Fram Strait is the primary gateway between the North Atlantic and the Arctic Ocean, where ongoing Atlantification is reshaping hydrographic conditions and marine ecosystems. We applied ML-PhyTAO, a machine learning based retrieval model for Arctic phytoplankton functional types (PFTs), to generate a consistent 28-year (1998–2025) satellite record of total chlorophyll-a (TChla) and major PFT chlorophyll-a concentrations (Chla) in the Fram Strait. Trained with an extensive Arctic phytoplankton pigment dataset and ocean color and environmental products, the model provides robust performance for long-term analyses. Using a dynamic water mass classification based on sea surface temperature and sea ice fraction, the Fram Strait was partitioned into Atlantic Water (AW), Mixed Water (MW), and Polar Water (PW). Distinct seasonal cycles were observed among water masses, with later but more intense phytoplankton blooms in AW. Peak biomass timing advanced by 2–13 days decade⁻¹ for TChla, diatoms, and haptophytes in both AW and MW. Trend analyses revealed declining biomass primarily in MW and the AW–MW transition zone, while haptophytes and dinoflagellates increased within AW. Sea surface salinity, temperature, and sea ice fraction showed stronger and more spatially coherent relationships with PFT variability than modeled nutrient fields, indicating that Atlantification related hydrographic changes are driving shifts in phytoplankton phenology and community composition. This study highlights the unique value of long-term, consistent PFT observations for detecting climate-driven ecosystem change, linking hydrographic variability to community shifts, and supporting future ecosystem and biogeochemical assessments and modelling studies in the rapidly changing Arctic Ocean.

Climate change has dramatically altered the Arctic seas with significant decrease in sea ice extent and thickness and warming water temperature. The ecological impacts of such change have been described for many parts of the Arctic Ocean, but long-term records of biological indicators are still missing. Among those, photosynthetic and accessory pigments are one of the key tools that aid quantification of phytoplankton and sea–ice algae biomass and characterisation of community composition. To address this gap, we present the first pan-Arctic compilation of in situ algal pigment data obtained exclusively by High-Performance Liquid Chromatography (HPLC), containing 10 798 samples collected across 77 Arctic research cruises between 2000 and 2024. As a result of large-scale collaborative effort, this database covers both open water and sea–ice environments across coastal, shelf and open domains. The database (https://doi.org/10.11583/DTU.29445104, Heidemann et al., 2026) includes measures of up to 26 pigments, with 8 major marker/accessory pigments being considered in this study, namely Alloxanthin (Allo), 19'-Butanoyloxyfucoxanthin (But-fuco), Chlorophyll a (Chl a ), Chlorophyll b (Chl b ), Fucoxanthin (Fuco), 19'-Hexanoyloxyfucoxanthin (Hex-fuco), Peridinin (Peri), and Zeaxanthin (Zea). This publicly available database provides crucial data that can be used to assess phytoplankton dynamics, validating remote sensing observations and can serve as a resource for future Arctic ecological- and modelling studies.

Understanding phytoplankton composition and distribution is essential for assessing marine ecosystem processes, yet satellite and in situ datasets each have substantial limitations. This study develops a fusion framework that combines gap-filled satellite-derived total chlorophyll a (TChla) and phytoplankton functional type (PFT) chlorophyll a (Chla) concentrations with in situ data points extracted from a hyperspectral flow-through spectrophotometric system (ACS). The objective is to generate spatially continuous and accurate Chla fields. The in situ hyperspectral absorption data obtained during the RV Polarstern PS113 expedition were used to derive TChla and PFT Chla concentrations with associated uncertainties, using an empirical orthogonal function based regression model (EOF-GLM). The ACS-derived estimates were then merged with the gap-filled satellite products through optimal interpolation to produce a consistent spatiotemporal dataset. The ACS-derived TChla and PFT Chla products showed strong agreement with the ground truth Chla concentrations (HPLC-Chla), with the coefficient of determination (R²) values above 0.66 for all groups and median percentage-deviation (MedPD) values between 27% and 46%. The fusion dataset further improved performance, reaching R² values higher than 0.69 and MedPD values between 21% and 44%. The combined product provides spatially and temporally continuous Chla fields and resolves regions with extensive satellite data gaps more effectively than either dataset alone. This fusion approach supports more reliable analyses of phytoplankton dynamics and can guide sampling strategies in areas with limited satellite coverage. Future work should focus on refining uncertainty estimates and testing the framework across different oceanic regions to improve its stability and broader applicability.

Climate change has dramatically altered the Arctic seas with significant decrease in sea ice extent and thickness and warming water temperature. The ecological impacts of such change have been described for many parts of the Arctic Ocean, but long-term records of biological indicators are still missing. Among those, photosynthetic and accessory pigments are one of the key tools that aid quantification of phytoplankton and sea-ice algae biomass and characterisation of community composition. To address this gap, we present the first pan-Arctic compilation of in situ algal pigment data obtained exclusively by High-Performance Liquid Chromatography (HPLC), containing 10.798 samples collected across 77 Arctic research cruises between 2000 and 2024. As a result of large-scale collaborative effort, this database covers both open water and sea-ice environments across coastal, shelf and open domains. The database (10.11583/DTU.29445104) includes measures of up to 26 pigments, with 8 major marker/accessory pigments being considered in this study, namely Alloxanthin (Allo), 19’-Butanoyloxyfucoxanthin (Butfuco), Chlorophyll a (Chl-a), Chlorophyll b (Chl-b), Fucoxanthin (Fuco), 19’-Hexanoyloxyfucoxanthin (Hexfuco), Peridinin (Peri), and Zeaxanthin (Zea). This publicly available database provides crucial data that can be used to assess phytoplankton dynamics, validating remote sensing observations and can serve as a resource for future Arctic ecological- and modelling studies.

Phytoplankton are vital to marine biogeochemical cycles and form the base of the marine food web. Comprehensive datasets offering a spatiotemporal perspective on phytoplankton composition are essential for assessing the impacts of climate change on marine ecosystems. Phytoplankton functional types (PFTs) classify phytoplankton based on their biogeochemical functions, enabling assessments of nutrient cycling, primary productivity, and ecosystem structure. However, satellite-derived ocean colour products like PFTs chlorophyll a (Chl a) concentrations are challenged by limited temporal and spatial coverage due to the exclusion of data collected under non-optimal observing conditions, such as strong sun glint, clouds, thick aerosols, straylight, and large viewing angles or due to the specific sensor configuration and sensor malfunction. This highlights the importance of gap-filling techniques for producing consistent datasets, which are currently missing for operational data sets. This study evaluates two robust gap-filling methods for satellite observations: Data Interpolating Empirical Orthogonal Functions (DINEOF) and Data Interpolating Convolutional Auto Encoder (DINCAE). These methods were applied to Sentinel 3A/B OLCI-derived Chl a concentration products in several regions of the Atlantic Ocean over three years of data, including total Chl a (TChl a) and Chl a concentration of five major PFTs, namely diatoms, dinoflagellates, haptophytes, green algae, and prokaryotic phytoplankton. The reconstructed datasets were assessed using test dataset evaluation and validated with in situ measurements collected during the transatlantic RV Polarstern expedition PS113 in 2018. The test dataset evaluation indicates that DINCAE performs slightly better than DINEOF in representing transient-scale variability, particularly within highly dynamic regions. DINCAE achieves an average root-mean-square-logarithmic-error (RMSLE) in cross-validation that is 66 % lower for TChl a and 16 % lower for PFTs compared to DINEOF. However, external validation using in situ measurements indicates better performance for DINEOF than DINCAE, with improved regression metrics for PFTs, including a 12.5 % better slope, 13.6 % better intercept, and 68 % higher coefficient of determination (R²). The gap-filled datasets exhibit slightly reduced but still robust accuracy compared to the original satellite data while preserving statistical trends, improving spatial structure restoration, and increasing matchup data for validation. It is concluded that DINCAE and DINEOF each have unique strengths for gap-filling ocean colour products. DINCAE performs well in complex water bodies, effectively reproducing patterns from the original satellite product. In contrast, DINEOF shows higher overall reliability, supported by independent validation, and is better suited for larger areas due to its lower computational demands.

Phytoplankton are vital to marine biogeochemical cycles and form the base of the marine food web. Comprehensive datasets offering a spatiotemporal perspective on phytoplankton composition are essential for assessing the impacts of climate change on marine ecosystems. Phytoplankton functional types (PFTs) classify phytoplankton based on their biogeochemical functions, enabling assessments of nutrient cycling, primary productivity, and ecosystem structure. However, satellite-derived ocean colour products like PFTs chlorophyll-a (Chla) concentrations are challenged by limited temporal and spatial coverage due to the exclusion of data collected under non-optimal observing conditions such as strong sun glint, clouds, thick aerosols, straylight, and large viewing angles or due to the specific sensor configuration and sensor malfunction. This highlights the importance of gap-filling techniques for producing consistent datasets, which are currently missing for operational data sets. This study evaluates two robust gap-filling methods for satellite observations: Data Interpolating Empirical Orthogonal Functions (DINEOF) and Data Interpolating Convolutional Auto Encoder (DINCAE). These methods were applied to Sentinel 3A/B OLCI-derived Chla concentration products in several regions of the Atlantic Ocean over three years of data, including total chlorophyll-a (TChla) and Chla concentration of five major PFTs, namely diatoms, dinoflagellates, haptophytes, green algae, and prokaryotic phytoplankton. The reconstructed datasets were assessed using test dataset evaluation and validated with in situ measurements collected during the transatlantic RV Polarstern expedition PS113 in 2018. The test dataset evaluation indicates that DINCAE outperforms DINEOF, particularly in capturing transient-scale features. DINCAE achieves an average root-mean-square-logarithmic-error (RMSLE) in cross-validation that is 66 % lower for TChla and 16 % lower for PFTs compared to DINEOF. However, external validation using in situ measurements indicates better performance for DINEOF than DINCAE, with improved regression metrics for PFTs, including a 12.5 % better slope, 13.6 % better intercept, and 68 % higher coefficient of determination (R²). The gap-filled datasets exhibit slightly reduced but still robust accuracy compared to the original satellite data while preserving statistical trends, improving spatial structure restoration, and increasing matchup data for validation. It is concluded that DINCAE and DINEOF each have unique strengths for gap-filling ocean colour products. DINCAE performs well in complex water bodies, effectively reproducing patterns from the original satellite product. In contrast, DINEOF shows higher overall reliability, supported by independent validation, and is better suited for larger areas due to its lower computational demands.

Phytoplankton are vital to marine biogeochemical cycles and form the base of the marine food web. Comprehensive datasets offering a spatiotemporal perspective on phytoplankton composition are essential for assessing the impacts of climate change on marine ecosystems. Phytoplankton functional types (PFTs) classify phytoplankton based on their biogeochemical functions, enabling assessments of nutrient cycling, primary productivity, and ecosystem structure. However, satellite-derived ocean colour products like PFTs chlorophyll-a (Chla) concentrations are challenged by limited temporal and spatial coverage due to the exclusion of data collected under non-optimal observing conditions such as strong sun glint, clouds, thick aerosols, straylight, and large viewing angles or due to the specific sensor configuration and sensor malfunction. This highlights the importance of gap-filling techniques for producing consistent datasets, which are currently missing for operational data sets. This study evaluates two robust gap-filling methods for satellite observations: Data Interpolating Empirical Orthogonal Functions (DINEOF) and Data Interpolating Convolutional Auto Encoder (DINCAE). These methods were applied to Sentinel 3A/B OLCI-derived Chla concentration products in several regions of the Atlantic Ocean over three years of data, including total chlorophyll-a (TChla) and Chla concentration of five major PFTs, namely diatoms, dinoflagellates, haptophytes, green algae, and prokaryotic phytoplankton. The reconstructed datasets were assessed using test dataset evaluation and validated with in situ measurements collected during the transatlantic RV Polarstern expedition PS113 in 2018. The test dataset evaluation indicates that DINCAE outperforms DINEOF, particularly in capturing transient-scale features. DINCAE achieves an average root-mean-square-logarithmic-error (RMSLE) in cross-validation that is 66 % lower for TChla and 16 % lower for PFTs compared to DINEOF. However, external validation using in situ measurements indicates better performance for DINEOF than DINCAE, with improved regression metrics for PFTs, including a 12.5 % better slope, 13.6 % better intercept, and 68 % higher coefficient of determination (R²). The gap-filled datasets exhibit slightly reduced but still robust accuracy compared to the original satellite data while preserving statistical trends, improving spatial structure restoration, and increasing matchup data for validation. It is concluded that DINCAE and DINEOF each have unique strengths for gap-filling ocean colour products. DINCAE performs well in complex water bodies, effectively reproducing patterns from the original satellite product. In contrast, DINEOF shows higher overall reliability, supported by independent validation, and is better suited for larger areas due to its lower computational demands.