<i>Enteromorpha</i> Coverage Information Extraction by 1D-CNN and Bi-LSTM Networks Considering Sample Balance From GOCI Images

Abstract

Satellite remote sensing (RS) with high temporal resolution and a large scale is the main method utilized for the dynamic monitoring of Enteromorpha prolifera (EP) blooms. Recently, deep learning (DL) methods have been applied to EP analysis due to their excellent feature representation. However, EP information extraction methods based on DL from low-spatial-resolution satellite images are still immature. The main problems with such methods include the insufficiency of spectral and spatial feature learning in low-resolution (LR) satellite images, as well as the sample imbalance that DL-based neural networks face in EP information extraction. To solve the above problems, a neural network-based EP extraction method considering sample balance is proposed in this paper and named EP rough-then-accurate extraction network (EPRTAE-Net). The method consists of two components: EP rough extraction (EPRE), a strategy that attends to sample balance, and EP accurate extraction (EPAE), a deep neural network based on one-dimensional convolutional neural network (1D-CNN) and bidirectional long short-term memory (Bi-LSTM), which fully considers the learned spectral information of each pixel and interpixel contextual dependencies. Geostationary Ocean Color Imager (GOCI) images with 500-m resolution were applied as the LR images in the experiments. The experimental results show that the proposed method has the capability to enhance adaptability in areas with different EP densities (achieving stable and excellent performance) and exhibits at least a 10% gain in F1-score and at least a 6% gain in IoU in extracting EP coverage information over other representative and traditional EP extraction methods in the Yellow Sea region.