Advancing harmful algal bloom detection with hyperspectral imaging: Correlation of algal organic matter and fouling indices based on deep learning

Abstract

The frequency and severity of harmful algal blooms (HABs) have been increasing due to the climate change. Algal organic matter (AOM), the primary contributor to HABs, causes membrane fouling in seawater reverse osmosis (SWRO) desalination plants. Water quality factors commonly used for SWRO operations during HAB events include the silt density index, modified fouling index, total organic carbon, transparent exopolymer particles, chlorophyll, and algae density. Traditional methods for measuring these factors are time-consuming and can disrupt decision-making processes. Therefore, continuous and real-time water quality monitoring using spectral sensors has become increasingly important. This study aimed to analyze AOM-based fouling- related water parameters using a hyperspectral imaging system and employed deep learning algorithms to simulate the fouling indicators. AOM, fouling indices, and hyperspectral images were used as input data, with convolutional neural network (CNN) and random forest (RF) models applied to extract band feature importance. The CNN (R2 = 0.71, mean squared error (MSE) = 435.21, mean relative error (MRE) = 23.46 %) outperformed the RF (R2 = 0.67, MSE = 2034.22, MRE = 25.76 %). The CNN model demonstrated an advantage in predicting fouling indices more consistently. Key spectral bands near 600 nm were identified for both models, which are crucial for detecting chlorophyll content, a strong indicator of algal bloom activity. Similarly, wavelengths above 730 nm were sensitive to organic matter, which is important for assessing AOM presence and fouling. These spectral ranges (604-686 nm for fouling and 733-876 nm for organic matter) are essential for monitoring fouling and bloom-related parameters. This study will provide more stable predictions that can enhance decision-making processes during HAB events.