Leveraging Sentinel-2 data to complement ground-based cloud cover statistics

Precise assessment of local cloud cover at a detailed spatial resolution can be useful for various types of applications, including for example forecast of power production by photovoltaic panels [1], crop quality prediction [2], or even micro-climatic and eco-dynamics studies [3]. Figure 1: The ACTRIS network of ground stations with cloud profiling capabilities. Various sensor networks exist, which locally sense the presence and profile of clouds; an example is reported in Figure 1. Such networks are however geographically too coarse to determine the cloud cover at any given location where it is needed with high positional accuracy as an input to an e.g. microclimatic model. Spaceborne remote sensing can be used as an alternate source of cloud cover information. As it is well-known, geostationary-orbiting (GEO) weather satellites such as Meteosat-8 can provide large-scale images of the Earth, from which cloud maps can be derived. Data from such platforms, however, is generated at a spatial resolution of a few kilometres, that may be too coarse for some applications, such as microclimate studies. An alternate approach, still based on spaceborne data but at a higher resolution, may be attempted using multispectral low-Earth-orbit sensors. As visible in Figure 2, sensors like Sentinel-2 provide a clear view of which areas are covered by clouds at a fairly high spatial resolution (10 m at best), and various algorithms are available in scientific literature to identify cloudy pixels, like Fmask [4], ACCA [5] and others [6]. Figure 2: two true-colour images derived from Sentinel-2 data over Dubai, UAE with (left) and without (right) cloud cover. Normalization of values for true colour representation results in altered tones on land. EO-data-processing environments such as the ESA Research and User Support (RUS) [7], or Google Earth Engine (GEE), feature indeed cloud detection and mapping capabilities, which we are exploiting for our work. In Figure 3, an example of output is visualized. In the case of Sentinel-2 data, a cloud mask is even provided as a standard component of the downloaded dataset. Figure 3: cloud masks automatically extracted from the datasets represented in figure 2. We have chosen 4 test sites around the globe, in 4 different climate zones according to the standard Köppen classification of climate zones [8], in order to diversify the contexts of our experiments:  Pavia, Italy (Cfa: hot-summer, temperate, humid climate)  Dubai, UAE (Bwh: very hot desert climate)  Nassau, Bahamas (Aw: savannah-like climate)  Stockholm, Sweden (Dwb: cold climate, dry winter) We have found and stored precise cloud cover statistics for all the 4 test sites. We then registered to ESA RUS, to handle and process Sentinel data, and to Google Earth Engine for LANDSAT data; we opted for remote, “cloud” processing systems to facilitate building and processing of thick stacks of data, in order to generate more significant statistics. We are now in the process of extracting long series of Sentinel-2 and LANDSAT data in order to systematically map cloud cover and generate the corresponding cloud statistics. These derived cloud statistics will be compared with the “ground truth” ones generated by ground stations and collected before stating the experiment. A thorough comparison will be made with the available, ground-based statistics. Although experiments are still at an initial stage, we have identified some issues that may interfere with the production of reliable statistics. One is illustrated in Figure 4, and consists of junk values along diagonal stripes in LANDSAT data, which happen to be mistaken for clouds by the cloud detection algorithm. Figure 4: the Dubai site. On the left, a true-colour representation of a LANDSAT multispectral dataset acquired in July 2017, on the right the corresponding cloud map. Note the diagonal spurious cloudy pixels matching the diagonal, bright line found in the upper half of the left-side image. The second one consists of sandy riverbeds that may be confused for clouds in summer months where water is at its minima. An example is shown in Figure 5, where the almost-dry Po river leaves visible large swathes of bright sand, which the cloud classifier wrongly associates with cloud presence. Figure 5: same as figure 4, on the test site of Pavia. Note the sandy riverbeds have been mistaken for clouds. Suitable countermeasures will be developed in order to suppress these causes of errors. Another possible matter is the temporal density of statistics. In order to assess the density of our statistical sample, we have made a preliminary search for LANDSAT data and found the figures illustrated in Table 1 for selected months in year 2017. Pavia (Italy) Dubai (UAE) Stockholm (Sweden) Nassau (Bahamas) January 4 4 none 2 April 5 4 6 2 July 6 4 7 2 November 2 4 none 2 Table 1: number of LANDSAT images identified on each test site for some selected months of year 2017 Some months are not well-covered, but these figures refer to a single sensor; we are confident that merging more datasets from multiple sources -as easily feasible within the selected processing environments- will remarkably improve the situation. Our results will be presented at the conference and discussed in light of the intended goal of this work. The purpose is not that of raising doubts on the validity of ground-based measurement, but rather to assess whether spaceborne data can be used as a valid replacement where ground stations may not be installed such as in remote locations whose statistics are nonetheless significant for purposes of climate studies. Future developments will include investigating possible fuzzy definitions of cloud cover and the introduction of multi-level statistics where the binary splitting into cloudy/non-cloudy class will be replaced by a "degree of cloudiness" on each image, and statistics adjusted accordingly. This work is being carried out as a group exercise within a Remote Sensing course at the University of Pavia, which has recently been selected as a new FabSpace under the H2020 “FabSpace 2.0” project of the European Union. A “Space Communication and Sensing” graduate track is currently active within the Engineering Faculty, and the aim of these exercises is that of showing the benefits of Earth Observation and encouraging public involvement in spaceborne monitoring of the terrestrial environment. References [1] Lipperheide, M., J. L. Bosch, and J. Kleissl. "Embedded nowcasting method using cloud speed persistence for a photovoltaic power plant." Solar Energy 112 (2015): 232-238. [2] Hoogenboom G. Contribution of agrometeorology to the simulation of crop production and its applications. Agricultural and Forest Meteorology. 2000;103(1–2):137–57. [3] Nicole M. Hughes, Kaylyn L. Carpenter, David K. Cook, Timothy S. Keidel, Charlene N. Miller, Junior L. Neal, Adriana Sanchez, William K. Smith, Effects of cumulus clouds on microclimate and shoot-level photosynthetic gas exchange in Picea engelmannii and Abies lasiocarpa at treeline, Medicine Bow Mountains, Wyoming, USA. Agricultural and Forest Meteorology, Volume 201, 2015, Pages 26-37, ISSN 0168-1923. [4] Z. Zhu, C.E. Woodcock: Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens. Environ., 118 (2012), pp. 83-94 [5] R.R. Irish, J.L. Barker, S.N. Goward, T. Arvidson: Characterization of the Landsat-7 ETM + automated cloud-cover assessment (ACCA) algorithm. Photogramm. Eng. Remote Sens., 72 (10) (2006), pp. 1179- 1188, 10.14358/PERS.72.10.1179 [6] Harb, Mostapha, Paolo Gamba, and Fabio Dell’Acqua. "Automatic delineation of clouds and their shadows in Landsat and CBERS (HRCC) data." IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 9, no. 4 (2016): 1532-1542. [7] European Space Agency (ESA) Research and User Support (RUS) service portal. [Online] Available at: https://rus-copernicus.eu/portal/ [8] Koppen, W. (1923). Die Klimate der Erde. Walter de Gruyter, Berlin, Germany (in German).