Enhancing the detection and classification of jellyfish in underwater environments by integrating spatial Bayesian deep neural networks for superior image quality and improved performance across various conditions

Species-level monitoring of jellyfish blooms is vital to reduce their impact on the environment and economy. This article proposes an underwater image classification system incorporating Image Enhancement, a Spatial Bayesian Neural Network (SBNN) and the Humboldt-Squid Optimization Algorithm (HSOA). There are 1417 raw images which have been labeled into 11 classes (10 Jellyfish Classes, 1 Fish Class), resulting in 3168 samples. These images were split into Training, Validation, and Testing datasets. The images were improved using the Risk-Sensitive Cubature Quadrature Kalman Filter (RSCQKF). This helps to reduce underwater haze and increase contrast in images. The improved images were used to train an SBNN classifier with HSOA to optimize network parameters. The proposed system achieved high accuracy (98.75%), precision (97.36%), and ROC-AUC (0.99). The results from ablation analysis proved that increasing the image quality and optimizing the parameters provided benefits. The proposed system produced good results for multi-class classification and can be used for automatic underwater jellyfish monitoring.