Cyanobacteria hot spot detection integrating remote sensing data with convolutional and Kolmogorov-Arnold networks

Prompt and accurate monitoring of cyanobacterial blooms is essential for public health management and understanding aquatic ecosystem dynamics. Remote sensing, in particular satellite observations, presents a good alternative for continuous monitoring. This study employs multispectral images from the Sentinel-2 constellation alongside ERA5-Land to enable broad-scale data acquisition. A simple deep convolutional neural network (CNN) architecture was proposed to analyze cyanobacteria (CB) concentration dynamics in Pigeon Lake, Canada, over five years. The model achieved an R 2 value of 0.81 and an RMSE score of 0.03 for the training set and 0.15 for the testing set, demonstrating high predictive accuracy. Using the Local Getis–Ord statistic, we identified and analyzed trends in hot and cold spots under the null hypothesis that such spots are randomly distributed, observing changes in their distribution and the median CB concentration in hot spots over time. Additionally, a Kolmogorov-Arnold Network (KAN) and dense neural networks (NN) with a single hidden layer were trained to classify sections of the lake shoreline into hot and no hot spots using the Dynamic World dataset within a 500 m radius of the lake. The KAN achieved a recall metric of 0.83 for detecting hot spots. • Satellite observations can provide continuous, broad-scale cyanobacteria monitoring. • A deep CNN model trained on field data predicts spatial cyanobacteria concentration. • Hot/cold spot analysis reveals spatiotemporal trends in cyanobacteria distribution. • We use a Kolmogorov-Arnold network to find hot spots based on land use patterns.