Cloud-Net: An End-To-End Cloud Detection Algorithm for Landsat 8 Imagery

Abstract

Cloud detection in satellite images is an important first-step in many remote sensing applications. This problem is more challenging when only a limited number of spectral bands are available. To address this problem, a deep learning-based algorithm is proposed in this paper. This algorithm consists of a Fully Convolutional Network (FCN) that is trained by multiple patches of Landsat 8 images. This network, which is called Cloud-Net, is capable of capturing global and local cloud features in an image using its convolutional blocks. Since the proposed method is an end-to-end solution no complicated pre-processing step is required. Our experimental results prove that the proposed method outperforms the state-ofthe-art method over a benchmark dataset by 8.7% in Jaccard Index.

Cross-Sensor Adversarial Domain Adaptation of Landsat-8 and Proba-V Images for Cloud Detection

“…Fully convolutional neural networks (FCNN) [15], most of them based on the U-Net architecture [16], produce very accurate results and have the advantage that they can be applied to images of arbitrary size with a fast inference time. Works such as Jeppesen et al [17], Mohajerani and Sahedi [18], [19], Li et al [20], or Yang et al [21] tackle cloud detection in Landsat-8 using Fully Convolutional Neural Networks trained in publicly available manually annotated datasets. They all show very high cloud detection accuracy outperforming the operational Landsat-8 cloud detection algorithm, FMask [22].…”

“…The SPARCS dataset, collected in the study of Hughes and Hayes [59], contains 80 1,000×1,000 patches from different Landsat-8 acquisitions. Finally, the 38-Clouds dataset of Mohajerani and Saeedi [18] has 38 full scenes mostly located in North America. Images and ground truth cloud masks from these datasets have been upscaled, to match Proba-V spectral and spatial properties, following the procedure described in section III-A.…”

Low-Power Neural Networks for Semantic Segmentation of Satellite Images

“…However, convolutional neural networks were not used. A lot of recent work has been done on cloud segmentation using CNN [25,26,28,5,22]. However, all these methods use computationnaly-heavy architectures on hyperspectral images and are designed to be used with powerful hardware on the ground, and thus, are not compatible with our use-case.…”

“…We use the Cloud-38 dataset, introduced by S. Mohajerani et al [26]. The dataset is composed of 4-band Landsat-8 images with a 30m resolution.…”

“…LeNetFCN architecture looks like the "C-FCN" architecture (see Fig.2) with F 1 = 3, F 2 = 5, except it uses 5x5 convolutions instead of 3x3 convolutions. We evaluate models using standard segmentation metrics: Total Accuracy, Precision, Recall, Specificity and Jaccard Index (IoU), as defined in [26].…”