| 要約: | The extraction of coral reef benthic material information is of great significance in coral reef remote sensing monitoring. Traditional information extraction methods for benthic organisms on coral reefs, such as Support Vector Machine(SVM) and the maximum likelihood method, have several drawbacks, including poor accuracy, low level of automation, and high time cost. At present, methods based on deep learning are being widely used in image segmentation, and satisfactory results have been obtained. Therefore, a segmentation network model based on improved U-Net and deep learning technology is designed to extract benthic material information from coral reefs. To improve segmentation details, a multi-input mode is set for each level of the encoder. The residual structure of ResNet34 is used as the encoder of the network to extract more abundant features. A new feature extraction block is designed by combining decomposition convolution, attention mechanism, and channel shuffle operation, to replace the common convolution layer in the encoder, bottom layer, and decoder; Concurrently, the attention mechanism is used to improve the far hop connection of the U-Net model, adjust the weight, and improve segmentation accuracy. Based on the GF-2 multispectral remote sensing image of Sanya, the extracted ground objects are classified as healthy coral reef, albino coral reef, algal mixture, sand, spray, deep sea area, and land. The dataset is established through visual interpretation and revision of an object-oriented method combined with Google Earth image. The experimental results show that mean Intersection over Union(mIoU)and average F1 score of the model in this study reached 67.17% and 78.7%, respectively. Compared with commonly used segmentation models, the proposed model performs better in visual effect and evaluation indicators. The ablation experimental results prove the effectiveness of the improved module.
|
|---|
