Ulva polifera Detection from High Resolution Remote Sensing Images Based on Dual-Path Convolutional Neural Networks

Objectives The green tide formed by Ulva prolifera (U. prolifera) is a harmful marine ecological disaster. The rapid and accurate detection is of great significance for timely management of U. prolifera and the healthy development of the marine industry.

Methods Because the boundary of U. prolifera area is difficult to be determined accurately in high resolution remote sensing images (HSRIs), an U. prolifera detection method for HSRIs based on dual-path convolutional neural networks (CNN) is proposed in this paper. First, a dual-path CNN semantic segmentation framework is designed based on the distribution characteristics of U. prolifera in HSRIs. The area and boundary of U. prolifera in HSRIs can be extracted simultaneously using the proposed framework. Then, the strategy for optimizing the initial U. prolifera area detection results based on U. prolifera boundary is proposed to improve the detection accuracy.

Results The experimental results show that the proposed method can extract U. prolifera accurately, with F1-score of 88.25%, intersection-over-union of 78.97% and over accuracy of 98.99%, which is better than other U. prolifera detection algorithms.

Conclusions The proposed method can obtain good results for the detection of different types of U. prolifera in HSRIs.