An AI-enabled high-resolution chlorophyll-a forecasting model that outperforms the operational global biogeochemical system
Authors: Yong Jiang, Zheng Jiang, Gaowei Zhang, Zengzhou Hao, Suixiang Shi, Chi Zhang, Wei Wang
Organizations: School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing, China . · Key Laboratory of Digital Ocean, National Marine Data and Information Service, Tianjin, China . · State Key Laboratory of Satellite Ocean Environment Dynamics, Second Institute of Oceanography, Ministry of Natural Resources, Hangzhou, China . · Shanghai Artificial Intelligence Laboratory, Shanghai, China .
Surface chlorophyll-a concentration is a key indicator of marine phytoplankton variability, and short-term forecasts of its distribution provide advance information for marine ecosystem monitoring and management. However, interactions between physical transport and biological processes complicate chlorophyll-a forecasting, and simulating these coupled processes with conventional numerical models imposes substantial computational demands. Here we introduce ChlCast, an AI model that generates daily forecasts of surface chlorophyll-a concentration over a 10-day horizon at approximately 4-km resolution between 60∘S and 60∘N. ChlCast combines factorized spatiotemporal convolutions with a multiscale architecture to learn from historical chlorophyll-a fields and climatological information. Compared with the operational 0.25∘ Copernicus global ocean biogeochemical forecasting system, ChlCast reduces mean absolute error and root mean square error by 70.74% and 47.63%, respectively, against satellite-derived Level-4 chlorophyll-a fields, and by 35.03% and 13.07% against independent BGC-Argo in situ observations. Coastal case studies further show that the 4-km forecasts retain coastal gradients and localized concentration maxima that are weakened by coarse-grid averaging. These results demonstrate the potential of observation-driven deep learning for high-resolution forecasting of surface chlorophyll-a variability at the near-global scale.
Figures & tables
Figure 1: Schematic of the ChlCast architecture. ChlCast uses log-space chlorophyll-a anomalies and the corresponding log-transformed daily climatology over the preceding 10 days as a two-channel input. Fixed temporal encoding provides information on the position of each input day. The model adopts a (2+1)D U-Net architecture, in which separate spatial and temporal convolutions capture spatiotemporal dependencies. The encoder (blue) extracts features at progressively coarser spatial scales, whereas the decoder (pink) restores spatial resolution and incorporates features from the corresponding encoder levels through skip connections. Numbers indicate the spatial dimensions of the feature maps. ChlCast generates log-space anomaly forecasts for the subsequent 10 days, which are combined with the corresponding target-day climatology and inverse-transformed to recover daily chlorophyll-a concentrations.
Figure 2: Comparison of ChlCast with the Copernicus operational global ocean biogeochemical forecasting system. a , Spatial distributions of MAE, RMSE, Pearson correlation coefficient ( r ) and mean bias for ChlCast and Copernicus, evaluated against satellite-derived L4 surface chlorophyll-a fields over 60∘S – 60∘N . b , Forecast performance against the satellite L4 reference fields as a function of lead time for ChlCast, Copernicus and three baselines (absolute persistence, fixed climatology and linear trend extrapolation). Metrics include MAE, RMSE, Pearson r and mean bias. c , Spatial distribution of the BGC-Argo profiles used for independent in situ validation. The validation dataset comprises 1,915 profiles from 505 floats. d , MAE, RMSE and Pearson r of ChlCast and Copernicus at forecast lead times of 1–10 days, evaluated against BGC-Argo surface chlorophyll-a observations. MAE, RMSE and mean bias are expressed in mgm−3 ; r is dimensionless.
Figure 3: Surface chlorophyll-a variability along the Israeli Mediterranean coast from 29 August to 2 September 2026. a , Satellite-derived L4 chlorophyll-a fields at 4-km resolution. b , ChlCast 1-day-ahead forecasts at 4-km resolution. c , The same ChlCast forecasts aggregated to a 0.25∘ Copernicus grid. d , Copernicus 1-day-ahead chlorophyll-a forecasts at 0.25∘ resolution. All panels use a common logarithmic colour scale for chlorophyll-a concentration ( mgm−3 ) to facilitate comparison of spatial structure and concentration gradients. Dates indicate forecast target dates.
Figure 4: Surface chlorophyll-a variability along the eastern coast of Hiroshima Prefecture and in the neighbouring Seto Inland Sea from 19 to 23 August 2026. a , Satellite-derived L4 chlorophyll-a fields at 4-km resolution. b , ChlCast 1-day-ahead forecasts at 4-km resolution. c , The same ChlCast forecasts aggregated to a 0.25∘ Copernicus grid. d , Copernicus 1-day-ahead chlorophyll-a forecasts at 0.25∘ resolution. All panels use a common logarithmic colour scale for chlorophyll-a concentration ( mgm−3 ) to facilitate comparison of spatial variability among neighbouring coastal waters. Dates indicate forecast target dates.
Architecture
Input configuration
MAE ↓
RMSE ↓
ACC ↑
(2+1)D U-Net
M0
0.0687
0.5127
0.4467
(2+1)D U-Net
M1
0.0677
0.5014
0.4675
(2+1)D U-Net
M2
0.0675
0.5004
0.4709
3D Swin-T U-Net
M0
0.0697
0.5147
0.4293
3D Swin-T U-Net
M1
0.0782
0.5021
0.4684
3D Swin-T U-Net
M2
0.0681
0.5000
0.4702
Table 1: Forecast performance across model architectures and input configurations (2023–2025). MAE and RMSE are expressed in mgm−3 . Arrows indicate whether lower ( ↓ ) or higher ( ↑ ) values represent better performance, and bold values denote the best result for each metric within each architecture.
Supplementary Fig. 1: Regional forecast performance relative to satellite-derived L4 chlorophyll fields. ChlCast and Copernicus forecasts are compared over 60∘ S– 60∘ N on a common 0.25∘ grid, using 56 available initializations from 14 August to 9 October 2026. a , Longhurst provinces. b , Geographical regions. c , Latitude bands. d , Annual-mean climatological chlorophyll classes ( mgm−3 ). Columns show lead days 1–10 and the pooled result (P). MAE and RMSE denote mean absolute error and root mean square error; their reductions are relative to Copernicus, with positive values indicating lower ChlCast errors. Δr is the ChlCast minus Copernicus Pearson correlation coefficient. Blue and red indicate positive and negative values; grey indicates unavailable data or undefined correlation. W is the percentage of valid sample area weight and Q is the percentage contribution to ChlCast weighted squared error, both pooled over days 1–10. Positive percentages below 0.01 are labelled <0.01 . Only sampled provinces and geographical regions are plotted. Unassigned and Unclassified denote unassigned regions and incomplete or missing climatological coverage, respectively. Region names and values are provided in Supplementary Table 1 .
Supplementary Fig. 2: Regional forecast performance relative to BGC-Argo chlorophyll observations. ChlCast and Copernicus forecasts are evaluated over 60∘ S– 60∘ N against biogeochemical Argo (BGC-Argo) observations from 15 August to 19 October 2026, using the same 16,180 equally weighted profile–forecast pairs. a , Longhurst provinces. b , Geographical regions. c , Latitude bands. d , Annual-mean climatological chlorophyll classes ( mgm−3 ). Columns show lead days 1–10 and the pooled result (P). MAE and RMSE denote mean absolute error and root mean square error; their reductions are relative to Copernicus, with positive values indicating lower ChlCast errors. Δr is the ChlCast minus Copernicus Pearson correlation coefficient. Blue and red indicate positive and negative values; grey indicates unavailable data or undefined correlation. Profiles and Floats give the numbers of distinct profiles and floats in the pooled samples; a profile can verify multiple lead times. Only sampled provinces and geographical regions are plotted. Unassigned and Unclassified denote unassigned regions and incomplete or missing climatological coverage, respectively. Region names, unsampled categories and full results are provided in Supplementary Table 2 .
Supplementary Fig. 3: Chlorophyll fields and 1-day-ahead forecasts in the southwestern Baltic Sea. The domain is 54∘ – 56∘ N, 12∘ – 16∘ E. Columns correspond to 24, 25, 26 and 27 August 2026. a , Satellite-derived L4 chlorophyll reference fields at 4-km resolution. b , 4-km ChlCast forecasts. c , The same ChlCast forecasts aggregated to a 0.25∘ grid. d , Copernicus forecasts at 0.25∘ resolution. Each forecast was initialized on the preceding day. All panels share a logarithmic colour scale of 1–30 mgm−3 ; values outside this range use the endpoint colours. The 4-km ChlCast fields and their 0.25∘ aggregates use the same valid marine display extent. Coarse-grid values are constant within each cell, without spatial interpolation or smoothing. Copernicus retains its own valid coverage; missing values are not filled. Grey denotes land or excluded locations, and pale blue denotes ocean locations without a valid displayed value.
Supplementary Table 1: Complete regional forecast evaluation against satellite-derived L4 chlorophyll fields. ChlCast and Copernicus results over 60∘ S– 60∘ N on the common 0.25∘ grid, for lead days 1–10 and their pooled result. The 56 available initialization dates span 14 August to 9 October 2026, with target dates from 15 August to 19 October. The electronic table contains four worksheets: Longhurst, Geographical, Latitude and Climatology. MAE, RMSE and bias are expressed in mgm−3 ; bias is forecast minus reference. MAE and RMSE reductions are relative to Copernicus, with positive values indicating lower ChlCast errors. r and Δr are dimensionless, and Δr is ChlCast minus Copernicus. Sample weight share is the fraction of valid area weight; Valid grid cells counts distinct evaluated cells. AE and SE contributions are the regional fractions of each model’s total absolute and squared error. AE reduction contribution is the regional contribution to the domain-wide reduction in absolute error. Plot ID corresponds to Supplementary Fig. 1 . Pooled 1–10 denotes the pooled result and ALL the full evaluation domain. Unassigned, Unclassified and unsampled categories are retained. Unavailable metrics are blank, with No data identifying unsampled categories. Bold rows identify pooled results only.
Supplementary Table 2: Complete regional forecast evaluation against BGC-Argo chlorophyll observations. ChlCast and Copernicus results over 60∘ S– 60∘ N for lead days 1–10 and their pooled result, using BGC-Argo observations from 15 August to 19 October 2026. The common sample comprises 16,180 equally weighted profile–forecast pairs from 1,915 distinct profiles and 505 floats. The electronic table contains Longhurst, Geographical, Latitude and Climatology worksheets. MAE, RMSE and bias are expressed in mgm−3 ; bias is forecast minus observation. MAE and RMSE reductions are relative to Copernicus, with positive values indicating lower ChlCast errors. Δr is the ChlCast minus Copernicus Pearson correlation coefficient. Matched pairs and Pair share give the pair count and its fraction of the full sample. Distinct profiles, Distinct floats, Target dates and Forecast cases report sample coverage. Profiles can verify multiple lead times, and distinct profile or float counts need not sum across regions. AE and SE contributions are the regional fractions of each model’s total absolute and squared error; AE reduction contribution is the regional contribution to the domain-wide reduction in absolute error. Plot ID corresponds to Supplementary Fig. 2 . ALL denotes the full domain. Unassigned, Unclassified and unsampled categories are retained. Unavailable metrics are blank; No data identifies unsampled categories. Bold rows identify pooled results only.
Architecture
Input
Bias
r
Log MAE
Log RMSE
MAE
RMSE
ACC
(2+1)D U-Net
M0
-0.0234
0.8284
0.1558
0.2373
0.0687
0.5127
0.4467
(2+1)D U-Net
M1
-0.0217
0.8329
0.1540
0.2339
0.0677
0.5014
0.4675
(2+1)D U-Net
M2
-0.0198
0.8328
0.1529
0.2323
0.0675
0.5004
0.4709
3D Swin-T U-Net
M0
-0.0209
0.8213
0.1571
0.2392
0.0697
0.5147
0.4293
3D Swin-T U-Net
M1
-0.0337
0.8325
0.2989
0.4520
0.0782
0.5021
0.4684
3D Swin-T U-Net
M2
-0.0198
0.8339
0.1553
0.2345
0.0681
0.5000
0.4702
Supplementary Table 3: Complete forecast evaluation across model architectures and input configurations. Results for (2+1)D U-Net and 3D Swin-T U-Net under input configurations M0, M1 and M2, evaluated over 60∘S – 60∘N for target dates from 1 January 2023 to 31 December 2025. The test set contains 1,087 complete 10-day windows. For each input configuration, all metrics are calculated from data pooled across forecast lead days 1–10. Bias, mean absolute error (MAE) and root mean square error (RMSE) are expressed in mgm−3 ; bias is forecast minus reference. r is the concentration-space Pearson correlation coefficient, and ACC is the anomaly correlation coefficient. Log MAE and Log RMSE are natural-log-space errors.
Marine biogeochemical forecasting is increasingly important for managing marine ecosystems and the carbon cycle, yet global, seasonal forecast products lag far behind physical oceanography, held back by the complexity of the processes involved and by data scarcity. We introduce BG4Sea, which to our knowledge is the first global, data-driven system to produce multivariate seasonal forecasts of the marine biogeochemical state. BG4Sea is a modular architecture with a column autoencoder that compresses the vertical column into a low-dimensional latent space, a latent forecaster propagates this representation forward in time, a surface-forcing conditioner that injects physical boundary information via Feature-wise Linear Modulation (FiLM), and a horizontal-coupling module that incorporates neighboring-column context through cross-attention. The model is trained and evaluated on the global ocean reanalysis BIORYS4 (NEMO/PISCES), and produces six-month forecasts at 1/4 degree, monthly resolution for dissolved chemistry, biology, and carbon-pool variables, outperforming persistence and climatology across most variables and lead times. We position BG4Sea as an interpretable baseline for future, more expressive approaches, and discuss predictability attribution to each component, alongside the model's structural limitations.
Phytoplankton play a central role in marine ecosystems and the global carbon cycle, with different groups contributing differently to ocean biogeochemical processes. While standard techniques exist for monitoring phytoplankton concentration from ocean-colour data, their community composition remains difficult to observe at large scales. Chlorophyll-a, widely available from satellite ocean-colour observations, is commonly used as a measure of phytoplankton biomass but provides limited information on taxonomic composition. Accessory pigments, some of which are diagnostic of important phytoplankton groups, offer additional information on community structure, but their retrieval from ocean-colour data is challenging because of limited spectral resolution and strong covariance with chlorophyll-a. In this study, we evaluate machine learning methods for estimating diagnostic pigment concentrations from multispectral satellite observations. Using a global dataset of 33,640 High Performance Liquid Chromatography (HPLC) measurements matched with ESA Ocean Colour Climate Change Initiative (OC-CCI) reflectance data, we compare Random Forest and TabPFN models trained on multispectral reflectance with baseline models using chlorophyll-a alone. A temporally stratified validation scheme is employed to reduce the effects of autocorrelation. Results show that multispectral models consistently outperform approaches based solely on satellite-derived chlorophyll-a, demonstrating that ocean-colour reflectance contains additional information relevant to pigment discrimination. Improvements vary by pigment, with those strongly correlated with chlorophyll-a showing limited gains, while others exhibit substantial improvement. These findings highlight the potential of machine learning to extract ecologically relevant information from satellite data beyond conventional chlorophyll-based approaches.
David Moffat, Angus Laurenson, Victor Martinez-Vicente +4
Plymouth Marine Laboratory, Plymouth, UK · National Center for Earth Observation, Plymouth Marine Laboratory, UK · Department of Earth and Environmental Sciences, Centre for Geography and Environmental Science, University of Exeter, Cornwall, United Kingdom
Global AI weather forecasting still relies mainly on uniform-resolution models, making it hard to combine regional refinement, two-way regional-global coupling, and affordable training cost. We introduce StretchCast, a global-regional AI forecasting framework built on a variable-resolution stretched cubed-sphere (SCS) mesh that preserves a closed global domain while concentrating resolution over a target region. Within this framework, we develop a one-step predictor, SCS_Base Model, and a rollout-oriented multistep predictor, SCS_FCST4 Model, to test the feasibility of SCS-based forecasting and the benefit of joint multistep training. Experiments use ERA5 with 69 variables over 1998-2022. Because training compute remains limited, this study uses a coarse-resolution proof-of-concept configuration rather than a final high-resolution system. Even with only about 7,776 effective global grid cells and roughly 0.875 degree resolution over the center-refined face, the 23M-parameter SCS_Base Model yields stable multivariate forecasts. With 83M parameters and training cost on the order of hours, SCS_FCST4 Model delivers competitive medium-range anomaly-correlation evolution over the target region after unified reprojection, especially for geopotential height, specific humidity, and part of the lower-tropospheric winds, while maintaining smooth cross-face continuity and realistic multiscale structure in typhoon and spectral analyses. These results support StretchCast as a practical lightweight foundation for global-regional AI weather forecasting.
Jin Feng
Institute of Urban Meteorology, China Meteorological Administration (IUM, CMA), Beijing, China · LeadSee Team, Beijing Meteorological Service