A three-step machine learning approach for algal bloom detection using stationary RGB camera images

Abstract

Stationary surveillance cameras deployed around lakes can provide continuous real-time observations of key water areas for harmful algal bloom (HAB). They can be used to supplement remote sensing-based monitoring in situations that satellites cannot handle. While some cameras were initially installed for other purposes, and the poses are not fixed during operation, hence, detecting HABs remains a challenging task due to the diverse surface features present in image frames. A novel three-step machine learning approach was proposed in this paper to address this problem. The acquired images are initially classified using the first model, and images with certain HABs undergo further examination. A second model is employed to generate a water mask, thereby eliminating interferences from non-water features. Finally, the third model is applied to detect and identify HABs specifically within water areas. The experiments showed that the three steps implemented in sequence can effectively extract distinct HABs from RGB images captured under various shooting poses. The overall pixel-level accuracy, intersection over union, and F1 score reached 0.83, 0.76, and 0.76, respectively, on 1969 images from August to September 2020. The novelty of our approach is attributed to that the combination of the three steps can significantly abate the adverse influence of an external environment; thus, the final detection can be performed with satisfactory accuracy. In practice, the approach was applied in Lake Chaohu and consistently reports the real-time status of HABs along the bank. It exhibits substantial potential for the application in eutrophic lakes to avoid HAB-induced secondary disasters.