Skip to main content
Back to timeline
arXivSource publication:

ChlCast forecasts chlorophyll-a at about 4 km: against the Copernicus operational system, errors fall by up to 70% on satellite reference fields

Synopsis

ChlCast is an observation-driven deep learning model that issues daily surface chlorophyll-a forecasts out to 10 days at roughly 4-km resolution between 60°S and 60°N. The operational Copernicus global ocean biogeochemical forecasting system previously ran on a 0.25° grid, where fine-scale structure is weakened by coarse-grid averaging. Against satellite Level-4 chlorophyll-a fields, ChlCast lowers mean absolute error and root mean square error by 70.74% and 47.63%; against independent BGC-Argo in situ observations, the two errors fall by 35.03% and 13.07%.

Source-provided article image: An AI-enabled high-resolution chlorophyll-a forecasting model that outperforms the operational global biogeochemical system
arXiv

Interpretation

Daily surface chlorophyll-a forecasts can now be produced at about 4-km resolution over a near-global domain from 60°S to 60°N, with a 10-day horizon. The prior operational global forecasts came from the Copernicus Global Ocean Biogeochemistry Analysis and Forecast system at 0.25° horizontal resolution, where narrow coastal gradients and localized concentration maxima are weakened by coarse-grid averaging. Across 56 valid forecast initializations between 14 August and 9 October 2026, with both ChlCast forecasts and satellite Level-4 reference fields aggregated to the Copernicus grid, pooled over the ten forecast days, mean absolute error and root mean square error were 70.74% and 47.63% lower than Copernicus, and the Pearson correlation coefficient was 0.8685 versus 0.3283.

The model learns the spatiotemporal evolution of surface chlorophyll-a from the previous 10 days of chlorophyll-a and daily climatological background, then extrapolates forward day by day for 10 days. Inputs are log-space anomalies relative to the daily climatological mean, with the corresponding log climatological mean supplied as an extra channel; this configuration gave the lowest mean absolute error and root mean square error and the highest anomaly correlation coefficient for both architectures. On the 2023–2025 test set, the (2+1)D U-Net with this input configuration reached a mean absolute error of 0.0675, root mean square error of 0.5004 and anomaly correlation coefficient of 0.4709; the 3D Swin-T U-Net with about 23.2 times as many parameters lowered root mean square error by only 0.08% while raising mean absolute error by 0.89%.

At native 4-km resolution the forecasts retain nearshore concentration gradients and localized concentration maxima that are visibly smoothed once aggregated to the 0.25° grid. When the same ChlCast forecasts were aggregated to the Copernicus grid, local chlorophyll-a maxima dropped and the share of within-cell variability in total spatial variance shifted. In the Israeli Mediterranean coast case from 29 August to 2 September 2026, daily spatial maxima in the aggregated forecasts were 32.78% to 50.47% lower than in the 4-km forecasts; within-cell variability accounted for 34.26% of total spatial variance in the 4-km ChlCast forecasts versus 41.27% in the satellite Level-4 reference fields.

Perspective

Evaluation covers daily 10-day forecasts on an approximately 4-km grid between 60°S and 60°N; the operational comparison is limited to 56 valid initializations from 14 August to 9 October 2026, and independent BGC-Argo validation uses 1,915 profiles from 505 floats between 15 August and 19 October 2026. The intended readers are researchers and engineers working on marine ecosystem forecasting and deep learning for the Earth system.

ChlCast underestimates chlorophyll-a concentrations at all forecast lead times, whereas Copernicus has an overall positive bias; classes with climatological concentrations of at least the highest threshold account for only 2.15% of total valid sample weight but contribute 56.79% of ChlCast's total weighted sum of squared errors. The satellite Level-4 product serves as both training target and primary gridded reference and carries interpolation and retrieval uncertainties in optically complex coastal waters, while BGC-Argo sampling is sparse in high-concentration waters. Differences from Copernicus also involve native resolution, modelling approach, information sources and initialization, so they cannot be attributed to a single factor.