Algoritmo para clasificación de familias de corales y su blanqueamiento

Abstract

This research presents an automated coral analysis system based on computer vision and deep learning techniques. The main objective is to classify coral families and evaluate bleaching severity from underwater images, integrating color restoration, segmentation, hierarchical classification, and health assessment into a single processing pipeline.

The system includes a color recovery module using the YCrCb color space and CLAHE (Contrast Limited Adaptive Histogram Equalization) to enhance luminance and correct the bluish-green cast common in underwater environments. A hybrid segmentation approach combining GroundingDINO and SAM (Segment Anything Model) ensures accurate coral isolation from complex marine backgrounds.

The hierarchical taxonomic classifier is built upon a multi-head ResNet-50 architecture, trained to simultaneously predict family, genus, and species levels, and optimized with the AdamW optimizer for improved convergence and generalization. A bleaching detection module analyzes luminance in the Lab space to produce severity maps and quantify bleached and at-risk regions.

Experimental results indicate a notable improvement in image quality (UIQM +27.3%, UIConM +36.4%) and an accuracy of 85.69% at the species level after integrating color recovery and segmentation. The system proved robust against preprocessing and illumination variations, validated with real-world underwater datasets collected under diverse environmental conditions.

In conclusion, the proposed end-to-end framework successfully transforms raw underwater imagery into quantitative biological information, supporting automated reef monitoring and contributing to the technological advancement of marine conservation tools.