Artificial intelligence will help predict harmful algal blooms

Algal blooms pose a serious threat to public health and aquatic ecosystems, as we have already seen in Poland as well. Published on June 3, 2025, as part of a special issue of New World: Advancing Water Applications Through Machine Learning and Artificial Intelligence, an article by Chinese researchers at Hohai University in Changzhou presents an interesting solution for predicting harmful algal blooms (HABs) in real time by integrating automated monitoring systems (AMS) with machine learning algorithms.

The goal of the study was to develop an energy-efficient predictive system for deployment on autonomous survey buoys.

Harmful algal blooms are on the rise

Algal blooms (HABs) pose a serious threat to public health and the balance of aquatic ecosystems. They are favored by excess nutrients (mainly nitrogen and phosphorus), and their development is strongly correlated with the physicochemical parameters of water – temperature, pH and electrical conductivity (EC). These phenomena are particularly exacerbated at temperatures above 20°C.

Existing monitoring methods, including satellite and manual monitoring, have their limitations – including low temporal resolution, interference of satellite images by cloud cover, and high operating costs. As an alternative, the study’s authors propose the BloomSense system. It is based on sensor buoys and an algorithm using four input variables: water temperature, pH, EC and chlorophyll-a (Chl-a) concentration.

What does BloomSense consist of?

The AMS system was deployed on buoys placed in the eutrophic freshwater reservoir As Conchas, located in southwestern Spain. Data were collected at 15-minute intervals over a period of 38 months. Monitoring took into account seasonal variability and hydrological conditions such as droughts and floods.

The predictive system is based on a complex hybrid machine learning model that includes:

• Random Forest (RF) – a.k.a. random forest, a method used to classify variables relevant to the prediction of Chl-a levels;
• SMOTE (Synthetic Minority Oversampling Technique) – a statistical technique that increases the number of cases in a data set in a balanced way;
• ResNet-18 convolutional neural network for spatial feature extraction from sensor data;
• LSTM (Long Short-Term Memory) – a recurrent neural network for modeling temporal dependencies;
• Softmax layer that predicts the probability of exceeding a critical Chl-a threshold.

The system also includes an alarm mechanism, activated when Chl-a concentrations exceed 10 µg/L – a value corresponding to Alert Level 1 according to WHO guidelines.

Key parameters for predicting blooms

The study used a data set that included means, standard deviations, minimum and maximum values for the variables temperature (°C), EC (μS/cm), pH and Chl-a (µg/L). Before analysis, the data were cleaned, normalized and transformed. By examining them, a strong correlation was identified between temperature, pH and Chl-a concentration, indicating their crucial importance in predicting harmful algal blooms.

An important aspect was to include the level of battery charge in the buoys as an indicator of system reliability – monitoring this parameter is expected to ensure uninterrupted system operation under conditions of limited energy resources.

Effectiveness of the proposed model

The model was tested on two data sets: from buoys located in stable inland waters and in dynamic coastal areas. The results show that the hybrid model (RF + SMOTE + ResNet-18 + LSTM):

• reduced the mean absolute error (MAE) by 26.2 percent compared to classical models;
• increased the F1-score (the harmonic mean of precision and sensitivity) by 70.2 percent in classifying the occurrence of a bloom;
• performed better in difficult, unstable conditions than the ResNet+LSTM model.

An in-depth analysis of the obtained measurements and prediction results showed that the model’s predictions matched well with the actual values of Chl-a concentrations, and exceedances of the 10 µg/l alarm threshold were accurately identified.

Cheaper and more effective prediction of harmful algal blooms

The authors developed a new hybrid machine learning model that combines deep feature extraction (ResNet-18) with sequential analysis (LSTM) while optimizing input variable selection and data balancing.

The study reached the following conclusions:

1. Continuous monitoring (every 15 minutes) with buoy sensors effectively replaces the limitations of satellite detection (including low resolution and lack of data when cloudy).
2. The use of low-cost input parameters (temperature, pH, EC, Chl-a) allows the solution to be scalable and useful in resource-constrained environments.
3. The integrated warning system allows rapid response to potential harmful algal blooms, supporting effective water resource management.

The study showed that it is possible to build an energy-efficient, low-cost and accurate algal bloom prediction system using sensors and machine learning algorithms. The model proved to be robust to environmental variability and effective under different hydrological conditions. In the future, the authors plan to analyze in detail the impact of battery charge level on data quality and overall performance of the proposed system.

Bibliography:

Rathore, W.U.A.; Ni, J.; Ke, C.; Xie, Y. BloomSense: Integrating Automated Buoy Systems and AI to Monitor and Predict Harmful Algal Blooms. Water 2025, 17, 1691. https://doi.org/10.3390/w17111691