Sea-land segmentation method based on an improved MA-Net for Gaofen-2 images

Abstract

This paper proposes EMA-Net, a fully convolutional neural network, to improve the effectiveness of sea-land segmentation on Gaofen-2 images. The aim is to address the issue of low segmentation accuracy in sea-land boundary regions when using remote sensing images for sea-land segmentation. The MA-Net network structure is enhanced by splitting the EfficientNet-B0 benchmark network into five convolutional blocks. The five downsampled convolutional blocks in MA-Net are then sequentially replaced. Furthermore, an extra loss term for the sea-land boundary region is incorporated through the introduction of a boundary region enhancement loss function. This approach encourages the network to focus on learning the boundary region between the sea and land. This improves the accuracy of its prediction. The study presents the results of segmentation experiments conducted on a constructed Gaofen-2 image dataset. The improved EMA-Net model, utilizing the boundary region enhancement loss, achieves better performance than other methods for both the overall region and the sea-land boundary region. The LR (Land Recall), LP (Land Precision), SR (Sea Recall), SP (Sea Precision), F1 Score (F1-Score), mIoU (Mean Intersection over Union), and EA (Edge Accuracy) are averaged over multiple experiments to reach 97.78%, 96.63%, 97.65%, 98.48%, 97.62%, 95.37%, and 87.08% respectively. Additional experiments on the IKONOS images also confirmed the adaptability of the proposed method to high-resolution imagery.