Automatic estimation of eelgrass cover using seafloor images

Abstract

This research assesses various methods to monitor eelgrass populations along Canada's eastern coasts. This is done by providing image data to deep learning models to estimate the percentage cover of eelgrass from these models. The dataset comprises ocean floor images obtained through kayak and diver surveys. Human-estimated percentage covers of eelgrass in these images are used to evaluate our models. Image classification and image segmentation (pixel-wise classification) approaches are evaluated in this research. Image classification determines the percentage cover estimates by discretizing these estimates into 6 classes representing eelgrass cover, and image segmentation does this by generating segmentation masks and extracting percentage cover information from the pixels identified as eelgrass. The models were either pre-trained on the BenthicNet dataset or used with random weight initialization, together with various pre-processing techniques for image segmentation. Two separate datasets were used to compare model performances on unseen data, where the first dataset corresponds to same-domain images because they are collected from locations nearby to the training data collection sites, ensuring similar characteristics, whereas the second dataset corresponds to different-domain images because they are collected from randomly distributed locations, providing a diverse set of characteristics. We found that the segmentation models underestimated eelgrass percentage cover, and the classification models overestimated. All the segmentation models used, performed equivalently, except the VGG16 encoder pre-trained on ImageNet dataset with a UNET decoder (V16I-UNet). All classification models had similar but worse results than the segmentation models.