GAN-Based Defogging and Multiscale Fusion Approach for UAV-Based Seagrass Bed Imagery Semantic Segmentation in Challenging Marine Environments

Seagrass beds, as one of China’s “Three Major Marine Ecological Systems,” have consistently been a focal point of monitoring efforts. However, the quality of UAV-captured aerial images depicting seagrass beds is often compromised by atmospheric conditions, particularly during foggy weather, resulting in a reduction in overall feature expression. This situation poses challenges such as weak feature representation and low contrast, making accurate identification of seagrass beds in complex marine environments a significant challenge. To address these challenges, this paper focuses on accurately identifying sea grass beds in complex marine environments. First, a DefoggingGAN network is proposed to mitigate the impact of foggy weather on the quality of seagrass bed images. This network, which is based on generative adversarial networks (GANs), is designed for image defogging. Second, for the precise identification of seagrass bed images captured by UAVs, a segmentation network for the Seagrass Bed Imagery Segmentation Network (SBISNet) is developed. This network incorporates an attention mechanism to capture context modules, focusing on high-resolution and low-resolution feature maps to obtain the global context. Additionally, a multiscale convolutional attention module is introduced to achieve the fusion of multiscale features. Furthermore, due to the limited availability of seagrass bed datasets in UAV scenarios, this paper utilizes UAVs to collect seagrass bed images and establishes a dataset named the Seagrass Bed Dataset. This research contributes to the broader exploration of cutting-edge technologies, particularly in the context of edge computing and the IoT, within the realm of UAV applications for environmental monitoring.