J2.2 Cloud Nowcasting on Satellite Images: A Novel Dataset and Experimental Comparisons

Monday, 13 January 2020: 11:00 AM
157AB (Boston Convention and Exhibition Center)
Andreas Holm Nielsen, Aarhus Univ., Aarhus, Denmark; and A. Wagner, A. Iosifidis, and H. Karstoft

Forecasting cloud development is a crucial part of modern weather forecasting systems, as clouds form an integral part
of Earth’s climate system and can cause uncertainty in weather prediction systems. Despite this, there is only a limited
amount of publicly available datasets for investigating cloud forecasting performance globally. This is likely due to the
difficulty of measuring and investigating clouds quantitatively, which is both complex and time-consuming, as clouds
are water droplets and ice crystals of a chaotic nature that exist in numerous layers of the atmosphere. Several different
methods for measuring clouds exist to date. Broadly speaking, these can be divided into surface-based methods, radar-
and lidar-based, satellite-based or combinatorial approaches [WMO/WWRP, 2012]. It can be argued that the most
consistent and objective long-term data source for clouds comes from geostationary satellites, as these can provide
high spatial- (1x1km to 3x3km) and temporal resolution (5- to 15-minute intervals) on a global scale in near real-time.
While geostationary methods might not be as accurate as some radar- and lidar-based profiling methods [Stephens et al.,
2018], they provide one critical ingredient for data-driven systems, which is an abundance of historical observations.
In this paper we describe and publish a novel satellite-based dataset dubbed "SatCloud", which contains 11 different
cloud types for multiple layers of the atmosphere. The cloud types are extrapolated using a segmentation algorithm
developed by [Le Gléau, 2019] as part of the Satellite Application Facility on Support to Nowcasting and Very Short
Range Forecasting (NWCSAF) software [EUMETSAT, 2018]. The algorithm uses multi-spectral satellite images in
addition to viewing geometry, geographical locations and atmospheric variables such as air temperatures and integrated
water vapour content. The dataset has a spatial resolution of 928 x 1530 pixels recorded with 15-min intervals for the
period 2017-2018 centered and projected over Europe, where each pixel represents an area of 3x3 km annotated with
the cloud type. An example observation of this dataset can be seen in Figure 1. This is the highest obtainable resolution
from the European Meteosat (MSG) satellites when using both visible- and infrared light [Müller et al., 2017]. To the
authors’ best knowledge, no such publicly available dataset exists, as current datasets come with either coarse spatial
resolution (9km to 31 km spatially) or low temporal granularity (one- to multiple hours between images) [NOAA,
nd, ECMWF, 2016]. The segmentation algorithm has been validated using SYNOP/SHIP observations for total cloud
cover and space-born CALIOP lidar for the major cloud classes we have described above [Le Gléau, 2019]. As a
supplement to the large dataset (400 GB), we also publish a low-resolution dataset of 15 x 15 km for Central Europe
designed for small-scale experiments and analysis.
For our second contribution in this paper, we present a baseline Deep Learning model applied to the small dataset
of SatCloud for nowcasting 16 frames into the future, which is equivalent to a prediction horizon of four hours. In
numerical weather prediction (NWP) systems, which are used worldwide as the primary source for weather forecasts,
numerical methods are applied to model the future relative to an initial state of the atmosphere using governing
equations from the field of physics [NOAA and National Weather Service, nd]. These atmospheric simulations are
both computationally and monetarily expensive. Our application focuses on nowcasting directly on the cloud-labelled
satellite images, which motivates the usage of deep learning as a potential alternative, where we instead learn the
physical spatiotemporal relationships in an end-to-end manner. This approach could also be used to complement
or replace specific parts of existing NWP models to free up computational resources or potentially improve overall
accuracy of weather forecasts.
Our model is based on pixel-wise classification using an autoencoder architecture for future frame prediction, which
enables us to optimize a) the encoding and decoding of spatial dimensions using 2D convolutional neural networks
(CNN) and b) capturing autoregressive correlations on the spatially-encoded representations using convolutional long
short-term memory (ConvLSTM) layers [Shi et al., 2015]. We note that similar approaches have shown great promise
in nowcasting precipitation using radar images [Shi et al., 2015, Shi et al., 2017], but the application to satellite images
for cloud forecasting is arguably more complex and has not been fully researched yet.
Our final contribution is an evaluation framework for measuring satellite-based cloud forecasting accuracy specifically
for our dataset. The evaluation scheme and the metrics used are based on best practices from the World Meteorological
Organization [WMO/WWRP, 2012]. To evaluate our model, we compare against two benchmark models; which are 1)
a simple persistence model and 2) historical forecast data from the global atmospheric ECMWF operational model.
We separate our training period to cover all of 2017 and the first six months of 2018 and our test period to cover the
remaining six months of 2018.
Preliminary results show great potential by our proposed Deep Learning model compared to both the persistence model
and the global ECMWF model (see Appendix B), and we are able to make predictions in near real-time on a much
finer spatial- and temporal scale than the ECMWF model. To generalize our results to applications outside nowcasting
for future work, we would have to include additional variables not directly observable in satellite images to model the
complex long-term interaction between clouds and other atmospheric variables. We also plan to evaluate our model
against local NWP models of higher resolution in the future, such as the German COSMO D-2 model [Baldauf et al.,
2018] or the French AROME model [Bouttier, 2007]. Finally, we note that SatCloud could be extended to include other
geostationary satellites, which provides the opportunity to develop global-scale weather-based forecasting systems with
multi-layer clouds as a central feature.
By publishing SatCloud we hope to advance the field of model-based and data-driven weather forecasting by having
cloud-labelled satellite images of high temporal and spatial resolution available to researchers and the weather forecasting
community worldwide. This should help the evaluation of sophisticated machine learning techniques and serve as
valuable input to existing global and local weather forecasting systems.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner