Foundation Department, Chongqing Medical and Pharmaceutical College, Chongqing, China.
School of Communication and Information Engineering, Chongqing University of Posts and Telecommunications, Chongqing, Chongqing, China.
PLoS One. 2023 Feb 6;18(2):e0280408. doi: 10.1371/journal.pone.0280408. eCollection 2023.
More than 50% of the images captured by optical satellites are covered by clouds, which reduces the available information in the images and seriously affects the subsequent applications of satellite images. Therefore, the identification and segmentation of cloud regions come to be one of the most important problems in current satellite image processing. Due to the complexity and variability of satellite images, especially when the ground is covered with snow, the boundary information of cloud regions is difficult to be accurately identified. The fast and accurate segmentation of cloud regions is a difficult point in the current research. We propose a lightweight convolutional neural network. Firstly, channel attention is used to optimize the effective information in the feature maps as a way to improve the network's ability to extract semantic information at each scale. Then, we fuse high and low-dimensional feature maps to enhance the network's ability to obtain small-scale semantic information. In addition, the feature aggregation module automatically adjusts the input multi-level feature weights to highlight the details of different features. Finally, we design the fully connected conditional random field to solve the problem that some noise in the input image and local minima during training is passed to the output layer resulting in the loss of edge features. Experimental results show that the proposed method achieves 0.9695 and 0.8218 for overall accuracy and recall, respectively, which has higher segmentation accuracy with the shortest time consumption compared with other state-of-the-art methods.
光学卫星拍摄的图像中超过 50%的图像被云层覆盖，这减少了图像中的可用信息，严重影响了卫星图像的后续应用。因此，云区的识别和分割成为当前卫星图像处理中最重要的问题之一。由于卫星图像的复杂性和可变性，特别是当地面被积雪覆盖时，云区的边界信息很难被准确识别。快速准确地分割云区是当前研究的一个难点。我们提出了一种轻量级卷积神经网络。首先，利用通道注意力来优化特征图中的有效信息，以提高网络在每个尺度上提取语义信息的能力。然后，我们融合高维和低维特征图，增强网络获取小尺度语义信息的能力。此外，特征聚合模块自动调整输入多层次特征的权重，突出不同特征的细节。最后，我们设计全连接条件随机场来解决输入图像中的一些噪声和训练过程中的局部最小值传递到输出层导致边缘特征丢失的问题。实验结果表明，与其他最先进的方法相比，所提出的方法的整体准确性和召回率分别达到 0.9695 和 0.8218，具有更高的分割精度和最短的时间消耗。
