CDUNet: Cloud Detection UNet for Remote Sensing Imagery

Abstract

Cloud detection is a key step in the preprocessing of optical satellite remote sensing images. In the existing literature, cloud detection methods are roughly divided into threshold methods and deep-learning methods. Most of the traditional threshold methods are based on the spectral characteristics of clouds, so it is easy to lose the spatial location information in the high-reflection area, resulting in misclassification. Besides, due to the lack of generalization, the traditional deep-learning network also easily loses the details and spatial information if it is directly applied to cloud detection. In order to solve these problems, we propose a deep-learning model, Cloud Detection UNet (CDUNet), for cloud detection. The characteristics of the network are that it can refine the division boundary of the cloud layer and capture its spatial position information. In the proposed model, we introduced a High-frequency Feature Extractor (HFE) and a Multiscale Convolution (MSC) to refine the cloud boundary and predict fragmented clouds. Moreover, in order to improve the accuracy of thin cloud detection, the Spatial Prior Self-Attention (SPSA) mechanism was introduced to establish the cloud spatial position information. Additionally, a dual-attention mechanism is proposed to reduce the proportion of redundant information in the model and improve the overall performance of the model. The experimental results showed that our model can cope with complex cloud cover scenes and has excellent performance on cloud datasets and SPARCS datasets. Its segmentation accuracy is better than the existing methods, which is of great significance for cloud-detection-related work.

Multiple Attention Mechanism Graph Convolution HAR Model Based on Coordination Theory

“…Due to the limitation of human body topology, it is difficult for the graph volume model to learn the relationship between various end nodes, which is often an important part of the action. In addition, the deep graph convolution model easily leads to the phenomenon of excessive smoothing of features [ 25 , 26 , 27 ], so it is not suitable to use the deep model [ 28 , 29 , 30 ]. Inspired by the dual attention network (DA-net) [ 31 , 32 ], an attention module is proposed.…”

Residual U-Net with Attention for Detecting Clouds in Satellite Imagery

“…Janocha and Czarnecki (2017) note that cross-entropy loss is often used for training neural networks without consideration for the alternatives. Indeed, cross-entropy has been the fiducial loss function in applications of deep learning to cloud detection (Hu et al, 2021;Guo et al, 2020). In the case of highly imbalanced images where the cloud pixels are few, however, this loss function can lead to simplistic solutions because many pixels can be wrongly classified as the background class to ensure a low loss value.…”

“…Following the results presented in the recent Cloud Mask Inter-comparison eXercise (Skakun et al, 2022) we test our model using existing L8 and S2 reference cloud datasets, which include the datasets of Hollstein (Hollstein et al, 2016), GSFC (Skakun et al, 2021), L8Biome (Foga et al, 2017), CESBIO (Baetens andHagolle, 2018), andPixBox (Paperin, 2021a,b). This allows us to evaluate the performance of our model using common points of comparison against which the most popularly employed cloud detection algorithms have also been evaluated (Zhu and Woodcock, 2012;Chen et al, 2021;Hu et al, 2021). For each exercise we compare the performance of our Attention ResUNet to the individual model(s) evaluated to have the best performance (in OA, BOA, PA, and UA) for each of these respective datasets.…”

“…The authors found ACD net outperforms other deep learning-based models in distinguishing such scenes, but does not cope well with thin cloud layers (Li et al, 2017;Zhan et al, 2017;Xia et al, 2019;Yang et al, 2019). Hu et al (2021) proposed a novel variation of U-Net they called CDUNet, that makes use of a ResNet-50 (He et al, 2016) as its encoder, which they then used to refine the edges of cloud segmentation regions. They performed their experiments on RGB bands of L8-SPARCS (256×256 pixels) and Google Earth's cloud and cloud shadow datasets (224×224 pixels).…”

An Overview on Visual SLAM: From Tradition to Semantic

“…The first step of semantic VSLAM is to extract semantic information from the images gained by the camera. Furthermore, semantic information based on image information can be achieved through classifying image information [205]. Traditional target detection relies on interpretable machine learning classifiers, such as decision trees and SVM, to classify and realize target features.…”