Â鶹ÒùÔº

A research team from the Hong Kong University of Science and Technology (HKUST) has developed a novel AI-powered tool named STIMP for diagnosing coastal ocean productivity and ecosystem health.

STIMP introduces a novel paradigm that imputes missing data and then predicts Chlorophyll-a (Chl-a) concentrations across large spatiotemporal scales. In tests across four representative global coastal regions, STIMP significantly outperformed existing geoscience tools, reducing the mean absolute error (MAE) for imputation by up to 81.39% and for prediction by 58.99%.

Accurate Chl-a prediction aids in early detection of harmful algal blooms, ecosystem protection, and provides data-driven insights for evidence-based policy-making. The research is in the journal Nature Communications, and the team was led by HKUST Professors Gan Jianping (Department of Ocean Science) and Professor Yang Can (Department of Mathematics).

Coastal oceans are the world's most productive marine ecosystems because terrestrial nutrient input and active hydrodynamics produce high biological productivity and biodiversity. The ecosystems of coastal oceans, however, are vulnerable to frequent and severe eutrophication, biogeochemical extremes, and hypoxia that substantially threaten the sustainability of these coastal environments and the blue economies of the coastal zones.

The concentration of Chl-a is a key indicator of the overall health of marine environments. Data-driven methods, utilizing remote sensing Chl-a data to achieve large-scale spatiotemporal diagnosis of ocean environmental quality, is a promising solution. However, there remain three challenges in developing large-scale spatiotemporal Chl-a data-driven prediction method.

First, temporal variations are difficult to capture in existing data. Second, spatial heterogeneity and relationships are difficult to model. Third, high rates of missing observations render spatiotemporal variations more challenging to comprehend.

To solve the above challenges, the HKUST research team developed an advanced AI-powered spatiotemporal imputation and prediction (STIMP) model for predicting Chl-a in the coastal ocean. STIMP decomposes the prediction of Chl-a into two sequential steps: 1) the imputation process, which reconstructs multiple potential complete spatiotemporal Chl-a distributions from partial observations, and 2) the prediction process, which accurately predicts Chl-a based on each reconstructed continuous and complete spatiotemporal Chl-a distribution.

Using Rubin's rules, the final Chl-a prediction is obtained by averaging the outcomes of multiple imputation and prediction processes. In this way, our STIMP method not only improves the overall predictive performance through accurate imputation of missing data but also provides confidence intervals to quantify the prediction uncertainties.

STIMP enables spatiotemporal imputation. STIMP significantly reduced MAE by 45.90~81.39% compared to the data interpolating empirical orthogonal function (DINEOF) method from geoscience, and by 8.92–43.04% compared to the state-of-the-art AI methods in four representative global coastal oceans. The Pearson correlation coefficients (PCC) of the STIMP imputed data and ground truth data were greater than 0.90 even when the rate of missing data equaled 90%.

With accurate imputation, STIMP offers spatiotemporal prediction, achieving MAE reductions of 58.99% over biogeophysical models and 6.54–13.68% over AI benchmarks. STIMP can be applied to global coastal oceans, offering a new approach for predicting oceans' Chl-a that typically have spatiotemporally limited data.

High-accuracy prediction of Chl-a concentration leads to significant benefits and solutions across various fields. In ecosystem management and conservation, it enables early detection of harmful algal blooms, allowing for proactive measures to protect aquaculture and coastal ecosystems. For policy and decision support, it provides data-driven insights for policymakers to design evidence-based fisheries regulations and pollution control measures.