Convolutional Neural Networks for On-Board Cloud Screening

Abstract

A cloud screening unit on a satellite platform for Earth observation can play an important role in optimizing communication resources by selecting images with interesting content while skipping those that are highly contaminated by clouds. In this study, we address the cloud screening problem by investigating an encoder–decoder convolutional neural network (CNN). CNNs usually employ millions of parameters to provide high accuracy; on the other hand, the satellite platform imposes hardware constraints on the processing unit. Hence, to allow an onboard implementation, we investigate experimentally several solutions to reduce the resource consumption by CNN while preserving its classification accuracy. We experimentally explore approaches such as halving the computation precision, using fewer spectral bands, reducing the input size, decreasing the number of network filters and also making use of shallower networks, with the constraint that the resulting CNN must have sufficiently small memory footprint to fit the memory of a low-power accelerator for embedded systems. The trade-off between the network performance and resource consumption has been studied over the publicly available SPARCS dataset. Finally, we show that the proposed network can be implemented on the satellite board while performing with reasonably high accuracy compared with the state-of-the-art.

Low-Power Neural Networks for Semantic Segmentation of Satellite Images

“…Our C-UNet matches the performance of our implementation of mobUNet [41], while using less memory thanks to the lack of skip connections. It also exceeds the performance of our implementation of StridedUNet [8], while being 20% smaller in terms of number of parameters. We see similar results on CloudPeru2 as on 38-Cloud.…”

“…Cloud segmentation using neural networks has already been integrated in an ARTSU CubeSat mission by Z. Zhang et al [41]. S. Ghassemi et al [8] have proposed a small strided U-Net architecture for onboard cloud segmentation. Their architectures are still too big for our use cases and are designed to work on 4-band images (RGB + NIR).…”

“…Their architectures are still too big for our use cases and are designed to work on 4-band images (RGB + NIR). Nevertheless, we include a 3-band implementation of MobUNet [41], MobDe-convNet [41] and "Plain+" strided U-Net [8] architectures in our comparison.…”

“…Zhang et al [41], are small versions of U-Net [31] and Deconv-Net [30] for onboard cloud segmentation, that use depth-wise separable convolutions [12], • StridedUNet, introduced by Ghassemi et al [8] for onboard cloud segmentation, which uses strided convolutions for downsampling, • LeNetFCN, an FCN variation of the well-known LeNet-5 architecture [17], which was originally created for written number classification. We replace the final Fully-Connected (Dense) layers by a 1x1 2D Convolution with a sigmoid activation, in order to output a segmentation map.…”

Transfer-learning-based cloud detection for Zhuhai-1 satellite hyperspectral imagery

“…The most important role of clouds in climate is to regulate the Earth's radiation balance, and they also play an important role in short-term weather forecasting and long-term climate change. Accurate distinction between cloud pixels and clear sky pixels to obtain high-precision cloud mask products is a basic requirement for extracting ground surface features using remote sensing data (Ghassemi and Magli, 2019); it also provides reliable data support for atmospheric and environmental applications by detecting the changes and movements of clouds over the atmosphere. The determination of high-precision clear sky pixels and cloud pixels is an important data support to expand remote sensing applications; therefore, cloud detection is a necessary part of remote sensing quantitative applications.…”