Raster Time Series: Learning and Processing
As the amount of remote sensing data is increasing at a high rate, due to great improvements in sensor technology, efficient processing capabilities are of utmost importance. Remote sensing data from satellites is crucial in many scientific domains, like biodiversity and climate research. Because we...
|Online Access:||PDF Full Text|
No Tags, Be the first to tag this record!
|Summary:||As the amount of remote sensing data is increasing at a high rate, due to great improvements in sensor technology, efficient processing capabilities are of utmost importance. Remote sensing data from satellites is crucial in many scientific domains, like biodiversity and climate research. Because weather and climate are of particular interest for almost all living organisms on earth, the efficient classification of clouds is one of the most important problems. Geostationary satellites such as Meteosat Second Generation (MSG) offer the only possibility to generate long-term cloud data sets with high spatial and temporal resolution. This work, therefore, addresses research problems on efficient and parallel processing of MSG data to enable new applications and insights.
First, we address the lack of a suitable processing chain to generate a long-term Fog and Low Stratus (FLS) time series. We present an efficient MSG data processing chain that processes multiple tasks simultaneously, and raster data in parallel using the Open Computing Language (OpenCL). The processing chain delivers a uniform FLS classification that combines day and night approaches in a single method. As a result, it is possible to calculate a year of FLS rasters quite easy.
The second topic presents the application of Convolutional Neural Networks (CNN) for cloud classification. Conventional approaches to cloud detection often only classify single pixels and ignore the fact that clouds are highly dynamic and spatially continuous entities. Therefore, we propose a new method based on deep learning. Using a CNN image segmentation architecture, the presented Cloud Segmentation CNN (CS-CNN) classifies all pixels of a scene simultaneously. We show that CS-CNN is capable of processing multispectral satellite data to identify continuous phenomena such as highly dynamic clouds. The proposed approach provides excellent results on MSG satellite data in terms of quality, robustness, and runtime, in comparison to Random Forest (RF), another widely used machine learning method.
Finally, we present the processing of raster time series with a system for Visualization, Transformation, and Analysis (VAT) of spatio-temporal data. It enables data-driven research with explorative workflows and uses time as an integral dimension. The combination of various raster and vector data time series enables new applications and insights. We present an application that combines weather information and aircraft trajectories to identify patterns in bad weather situations.|
|Physical Description:||199 Pages|