AI models crack the code of algal blooms in India’s Ashtamudi estuary

On the south-west coast of India, where the Kallada River meets the Arabian Sea through a perennial opening, lies the Ashtamudi estuary, the deepest estuary in the state of Kerala and one of the country’s most treasured Ramsar wetland sites. Like many productive estuaries around the world, Ashtamudi is under pressure. Agricultural fertilizers washed from nearby catchments, waste water discharge, cage fish farming and unscientific tourism practices have pushed the estuary into a eutrophic state, meaning it carries an excessive load of nutrients such as nitrogen and phosphorous. These nutrients fuel algal blooms, and the single most telling signature of those blooms is chlorophyll-a, the green pigment at the heart of photosynthesis in phytoplankton. A new study published in Estuarine Management and Technologies has now shown that artificial intelligence models, when fed the right ingredients, can predict chlorophyll-a dynamics in this estuary with astonishing accuracy, approaching a coefficient of determination of nearly one.

Chlorophyll-a matters far beyond the laboratory. It serves as a proxy for eutrophication levels and as a measure of overall phytoplankton biomass, which in turn represents the primary productivity of the ecosystem. When chlorophyll-a concentrations climb too high, the consequences ripple outward: bad odour and taste in the water, depletion of dissolved oxygen, enhanced turbidity and organic load, and ultimately damage to fisheries, tourism and public health. Monitoring agencies would ideally track chlorophyll-a daily, but laboratory determination is time-consuming, so most agencies settle for monthly sampling. Prediction models offer a way to fill those gaps, yet chlorophyll-a behaves non-linearly with its driving factors, and the factors themselves are site-specific. Models built for lakes or coastal seas rarely transfer cleanly to estuaries, where freshwater and seawater mix in constantly shifting proportions controlled by river flow and tidal flow.

That dynamic mixing is precisely what previous models have overlooked. The research team, working with monthly water quality data collected by government agencies between 2012 and 2018 at a mid-estuarine station, recognized that salinity offers an easily accessible, indirect measure of estuarine mixing intensity. When river discharge is high during the monsoon, salinity in the estuary can drop to as little as 1.25 parts per thousand; during the pre-monsoon, when freshwater influx is negligible, salinity can reach 30 parts per thousand. No earlier chlorophyll-a prediction study had accounted for salinity as an input variable in an estuarine setting, so the team designed two modelling scenarios: one using turbidity, total nitrogen and total phosphorous alone, and another adding salinity to the mix.

The raw data, covering pH, turbidity, salinity, dissolved oxygen, total nitrogen, total phosphorous and chlorophyll-a, were cleaned, checked for outliers and normalized to enhance the performance of the AI models, then restored to their original form after network training. Correlation analysis revealed which parameters truly drive chlorophyll-a in Ashtamudi. Total nitrogen and total phosphorous showed the strongest correlations, significant at the 99% confidence level, followed by turbidity and salinity, significant at the 95% level. Dissolved oxygen had a weak negative correlation and pH a weak positive one, but neither was statistically significant, so those two were excluded. Nitrogen and phosphorous act as causative parameters of chlorophyll-a, while turbidity behaves as a response variable, since higher algal concentrations themselves enhance water turbidity.

With the inputs settled, the researchers trained and tested seven models, splitting the data 70% for training and 30% for testing and validation. Three belonged to the artificial neural network family: the Back Propagation Neural Network, the workhorse of environmental modelling, which minimizes prediction error by adjusting weights through the back propagation algorithm across input, hidden and output layers; the Generalized Regression Neural Network, a radial basis function network prized for fast learning and rapid convergence on non-linear problems; and the Recurrent Neural Network, a deep learning architecture that works through parameter sharing and state transformation. Four more models came from the Adaptive Neuro Fuzzy Inference System, or ANFIS, a hybrid that fuses a fuzzy Sugeno system with a multilayer feed-forward network across six layers, from input through fuzzification, rules, normalization and defuzzification to output. In ANFIS, premise parameters are tuned by back propagation while consequent parameters are adjusted by least squares, and the choice of membership function, whether triangular, trapezoidal, Generalized Bell or Gaussian, proves crucial to performance.

The results were striking. Adding salinity improved predictability across the board, but the Generalized Regression Neural Network delivered the standout performance. For the scenario incorporating salinity, GRNN achieved a coefficient of determination of 0.9976 during training and 0.9982 during testing, with a root-mean-square error of just 0.01, nearly a perfect fit between predicted and observed chlorophyll-a. By contrast, the Recurrent Neural Network was the worst performer for the salinity-free scenario, with coefficients of determination of only 0.37 in training and 0.21 in testing, before improving drastically once salinity was included. The Back Propagation Neural Network fared poorly during testing without salinity, with a root-mean-square error as high as 0.85, but cut that error by 90.6% when salinity entered the model. These contrasts highlight that BPNN and RNN are highly sensitive to input parameters, while GRNN remains stable across varying salinity conditions.

Among the ANFIS models, those built with Gaussian and Generalized Bell membership functions outperformed the triangular and trapezoidal versions, particularly for the salinity scenario, reaching coefficients of determination of 0.98 during testing. Validation plots told a consistent story: without salinity, the neural networks over-predicted chlorophyll-a at low concentrations and under-predicted at high ones, but the inclusion of salinity largely erased those biases. The overall range of model performance across the study spanned coefficients of determination from 0.73 to 0.99, confirming that models incorporating salinity alongside total nitrogen, total phosphorous and turbidity well-predicted chlorophyll-a concentration in this estuarine environment.

Beyond model rankings, the study uncovered a fascinating ecological pattern in how salinity shapes chlorophyll-a itself. Below salinity values of 5 parts per thousand, chlorophyll-a showed an overall decreasing trend with salinity, but the relation was not strongly significant. Above that threshold, between 5 and 30 parts per thousand, chlorophyll-a rose pronouncedly with increasing salinity. The team offers a physical explanation: at lower salinity levels, appreciable freshwater flow creates salinity gradients that render estuarine reaches partially stratified or stratified, limiting mixing. Higher near-surface salinity signals a well-mixed water column, and that enhanced mixing promotes the penetration of light into deeper zones, boosting primary productivity and, with it, chlorophyll-a concentration. In other words, estuarine mixing itself enhances primary productivity through increased light penetration, and salinity serves as an indirect representation of the mixing conditions that play a crucial role in chlorophyll-a dynamics.

The implications extend well beyond one Indian estuary. Because salinity is an easily accessible parameter routinely measured by monitoring agencies, incorporating it into chlorophyll-a prediction models offers a replicable template for estuarine management worldwide. The authors emphasize that predictive models are important tools that fit well within estuarine environments due to their replicability, and that the salinity effect needs testing in other estuaries and with other modelling approaches to validate and generalize the findings. Future work should extend the analysis across space and time, examine performance under various tidal conditions, and probe the influence of other water quality parameters on chlorophyll-a prediction. For now, the message is clear: in the churning, shifting world of an estuary, the key to forecasting algal blooms may lie in a single, humble measurement, the saltiness of the water itself.

Subject of Research: Predicting chlorophyll-a dynamics in the Ashtamudi estuary using ANN and ANFIS techniques

Article Title: Predicting chlorophyll-a dynamics in the Ashtamudi estuary using ANN and ANFIS techniques

Article References: Priya, K. L., Haddout, S., & Casila, J. C. C. (2024). Predicting chlorophyll-a dynamics in the Ashtamudi estuary using ANN and ANFIS techniques. Estuarine Management and Technologies, 1, 57-68. https://doi.org/10.3897/emt.1.130278

Image Credits: AI Generated

DOI: 10.3897/emt.1.130278

Keywords: chlorophyll-a, Ashtamudi estuary, ANN, ANFIS, eutrophication, salinity, machine learning, water quality, Predicting, dynamics, Ashtamudi, estuary