基于深度学习的青海湖裸鲤产卵场遥感识别方法

鱼类产卵场会随着外部环境条件的改变而发生变化, 因此, 快速、有效地定位鱼类产卵场对于开展水生生物资源调查、珍稀水生动物保护等工作具有重要的意义。本研究基于无人机航拍影像和青海湖裸鲤(Gymnocypris przewalskii)产卵场实地调查结果, 构建了深度学习模型, 以分析将深度学习模型应用于青海湖裸鲤产卵场识别中的可行性。模型训练交并比精度和像素精度分别为 0.870 和 0.996, 验证交并比精度和像素精度分别为 0.648 和 0.985, 虽然精度低于一般的遥感影像或图像分割精度, 但从测试的结果来看, 深度学习模型可以识别到约 79%的产卵场, 但尚不能精确地分割出产卵场, 可以作为一种辅助手段, 应用到青海湖裸鲤产卵场的识别中。

Remote sensing recognition of spawning grounds of Gymnocypris przewalskii based on deep learning

Fish spawning grounds are affected by changes in environmental conditions. Rapid and effective identification of fish spawning grounds is crucial for assessing aquatic resources and protecting rare aquatic animals. In this study, based on aerial images captured by drones and field surveys, we identified 77 spawning grounds of Qinghai Lake naked carp (Gymnocypris przewalskii). By constructing a sampling method suitable to rivers and improving sampling technology, we obtained 807 sets of training samples, 56 sets of verification samples, and 23 sets of test samples; of these, the test samples were not enhanced. By linking the lightweight UNet and NestedUNet models in a chain, a UNet-NestedUNet deep learning model was established, and its performance was compared with UNet and NestedUNet models with twice the number of channels. The results showed that UNet-NestedUNet model performed better on the validation set. Model validation accuracy was the highest in the 51st step of training. The training intersections over union ratio and accuracy of the model were 0.870 and 0.996, respectively, and the validation intersection over union and accuracy were 0.648 and 0.985, respectively. Considering the test results, accuracy was lower than general remote sensing image or image segmentation accuracy. However, the deep learning model effectively identified most of the spawning grounds of Qinghai Lake naked carp (79%). In the entire area, the model identified 61 out of 77 spawning grounds, and only one spawning ground in the test area was not identified. In addition, the model identified a large number of unmarked areas as spawning grounds; these may be small areas not considered during manual identification, or misjudged by the model. This discrepancy may be caused by the relatively small number of actual samples used for training and the uncertainty of the spawning ground itself. The former can be resolved by obtaining more spawning ground samples of Qinghai Lake naked carp through long-term data accumulation. For the latter, it is necessary to further refine spawning field boundaries. At present, deep learning can be used as an auxiliary means to identify spawning grounds of Qinghai Lake naked carp. With the continuous increase in the number of cumulative samples in the future, the performance of the model will further improve. Therefore, the deep learning model has prospective applications in the identification of fish spawning grounds