Marine Disaster Identification Based on Fusion of Multi-Scale and Inverted Residuals Encoder

Achieving precise and high-efficiency detection of harmful algal blooms (HABs) serves as a pivotal cornerstone in developing red tide early-warning frameworks. Yet, owing to the striking resemblance in the cellular morphological traits of HABs, prevailing image categorization networks confront considerable hurdles when it comes to the accurate differentiation of such algal blooms. To tackle this predicament, the present research introduces a deep learning framework (FDIR) tailored for marine hazard identification, which hinges on multi-scale feature merging and inverted residual encoders, with the goal of boosting the precision of HABs classification. This framework is structured with a dual-branch design: the first branch is anchored by the inverted residual encoder module, primarily tasked with extracting and depicting the global semantic characteristics of images; the second branch, through the seamless integration of convolutional neural networks and a dual-channel attention mechanism, realizes the accurate acquisition and amplification of critical feature data related to HABs. Contrastive trials were carried out utilizing the public HAB dataset provided by AICO Labs. The outcomes indicate that, in comparison with prevalent classification network frameworks, the FDIR framework reaches a top-tier classification accuracy of 90.22%, and also demonstrates superior competitive capabilities in aspects such as model training duration and convergence efficacy.