Integrated monitoring of icing-up of selected swiss lakes

Various lake observables, including lake ice, are related to climate and climate change and provide a good opportunity for long-term monitoring. Lakes (and as part of them lake ice) is therefore considered an Essential Climate Variable (ECV) (WMO, 2018) of the Global Climate Observing System (GCOS). In Switzerland, the implementation of GCOS is coordinated by the Swiss GCOS Office at the Federal Office for Meteorology and Climatology MeteoSwiss. In 2007 (updated in 2018), MeteoSwiss published the first national inventory of the most important climate observations in Switzerland (MeteoSwiss, 2018). For each ECV, including observations of lake ice subsumed under the ECV “Lakes”, and international centre, the inventory report identified possible gaps regarding the legal basis, definition of responsibilities, and availability of financial resources for the continuation of observations and operation, respectively. In 2008, considering the findings of this report, the Federal Council approved a financial contribution to secure the continuation of time series and international centres at risk of discontinuation, as a long-term contribution to GCOS. Concerning the ECV “Lakes”, in the global National Snow and Ice Data Center (NSIDC), Boulder, Colorado database on lake ice-on/off dates, currently only Lake St. Moritz is partly included (till 2012). Further, existing observations and data from local authorities and publications are not systematic and come from different, uncoordinated and not secured sources. Traditionally, on-shore observers collected the information of lake ice recording the visible frozen-edge. Within the past two decades due to lack of budget and/or human resources, the number of field observations declined and mostly totally stopped. At the same time, the potential of different remote sensing sensors covering varying time periods and spatial coverage to measure the occurrence of lake ice was demonstrated by several investigations. Thus, following the need for an integrated multi-temporal monitoring of lake ice in Switzerland, MeteoSwiss in the framework of GCOS Switzerland supported this 2-year project to explore not only the use of satellite images but also the possibilities of Webcams and in-situ measurements. The aim of this project is to monitor some target lakes and detect the extent of ice and especially the ice-on/off dates, with focus on the integration of various input data and processing methods. Regarding ice-on/off dates, the GCOS requirements are daily observations and an accuracy of +/- 2 days. The target lakes were: St. Moritz, Silvaplana, Sils, Sihl, Greifen and Aegeri, whereby only the first four were mainly frozen during the observation period and thus processed. These lakes have variable area (very small to middle-sized), altitude (low to high) and surrounding topography (flat/hilly to mountainous) and freeze generally often, covering thus difficult to medium difficulty cases, regarding area, altitude and topography. For all image data, only cloud-free pixels lying entirely within the lake were processed. The observation period was mainly the winter 2016-17. During the project, various approaches were developed, implemented, tested and compared. Firstly, low spatial resolution (250 - 1000 m) but high temporal resolution (1 day) satellite images from the optical sensors MODIS and VIIRS were used. Secondly, and as a pilot project, the use of existing public Webcams was investigated for (a) validation of results from satellite data, and (b) independent estimation of lake ice, especially for small lakes like St. Moritz, that could not be possibly monitored in the satellite images. Thirdly, in-situ measurements were made in order to characterize the development of the temperature profiles and partly pressure before freezing and under the ice-cover until melting. Besides the validation of the results from other data, this in-situ data is used to calibrate a one-dimensional physical lake model so that the criteria for freezing of different Swiss lakes can be derived as a function of meteorological and morphometric conditions. It is expected that the developed methods and software can be used, possibly with some modifications, also for other data acquired for the same lakes or other ones in Switzerland and abroad. The project is a feasibility study, which should lead to a comparison and analysis of the above three techniques and recommendations to MeteoSwiss for further actions. This report presents the results of the project work. Using MODIS and VIIRS satellite data, ETHZ proposed a processing chain for lake ice monitoring. We tackled lake ice detection as a pixel-wise, two-class (frozen/non-frozen) semantic segmentation problem with Support Vector Machine (SVM) classification. Four different lakes were analysed: Sihl, Sils, Silvaplana and St. Moritz using both MODIS and VIIRS data. While we have concentrated on lakes in Switzerland, the ETHZ methodology is generic and the results could hopefully be directly applied to other lakes with similar conditions, in Switzerland and abroad, and possibly to similar sensors. To assess the performance, the data from both MODIS and VIIRS were processed from one full winter (2016/17) including the relatively short but challenging freezing and melting periods, where frozen and non-frozen pixels co-exist on the same lake. Using MODIS data, we also processed dates from 2011/12 and we demonstrate that the ETHZ approach gives consistent results over multiple winters, and that it generalizes fairly well from one winter to another. For both MODIS and VIIRS, we have also shown that the model generalizes well across lakes. ETHZ dealt with the feasibility of lake ice monitoring using Webcam images as a supplement or alternative to other monitoring methods. Therefore, publically available Webcam images capturing six Swiss lakes were downloaded from the Internet in the periods of the winters 2016/2017 and 2017/2018. Our evaluation was based on two cameras (high- and low-resolution) monitoring the lake of St. Moritz. To predict the lake ice coverage of the observed water body, fully Connected Neuronal Networks (part of Deep Learning and Artificial Intelligence methods) were utilized. Given an input image, such networks are designed to predict pixel-wise class probabilities. For the problem at hand, the target classes were: snow, ice, water and clutter. Several tests were conducted to identify important parameters for the intensive training of such networks. Moreover, generalization capabilities with respect to differing cameras and to data recorded in different winters were investigated. In contrast to optical satellite data, Webcams typically record multiple images per day. A median-based strategy to fuse such results to derive daily predictions was implemented. Moreover, possibilities of further exploiting temporal redundancy to improve predictions were explored by using two additional network architectures. The Remote Sensing Research Group, Institute of Geography, University of Bern focuses on the development of a physical approach to monitor lake ice based on data of the satellite sensor VIIRS (Visible Infrared Imaging Radiometer Suite) on-board of NOAA satellites Suomi-NPP. The advantage of the VIIRS data is the daily temporal resolution and improved spatial resolution of 375m considering the high-resolution channels (I-channels). A comprehensive pre-processing chain has been developed and implemented to efficiently obtain projected data of the region of interests (target lakes in Switzerland) from multiple scans. Not only the scientific data records of VIIRS data (e.g., radiometric data, geolocation information) were processed but also environmental data records (e.g., VIIRS M-band surface temperature), and intermediate products (e.g., cloud masks), which were needed for product retrieval. The developed approach considers surface temperature and reflectance values in the visible and near-infrared spectra. The method relies on the assumption that a frozen water surface has a temperature below 0°C and an ice layer increases the reflectance in the visible/near-infrared wavelengths. Retrieval of surface temperature using only one channel in the atmospheric window requires a procedure to correct for atmospheric effects (attenuation). Therefore, a single channel PMW (Physical Mono Window) model has been adapted to the thermal I-band data of the VIIRS sensor (I05) considering atmospheric data from the European Center for Medium Weather Forecast (ECMWF), which were required for the radiative transfer modelling to minimize atmospheric attenuation. The accuracy of the retrieved VIIRS I-band PMW Lake Surface Water Temperature (LSWT) has been assessed with cross-satellite comparison and temperature based validation for Lake Geneva (station Buchillon) and the target lakes Lake Greifen and Lake Sils. The assessment of thermal infrared-derived surface temperatures indicates an overall good performance of the physical mono window LSWT retrieval method for VIIRS I-band data. Therefore, lake ice detection, based on LSWT, the normalized difference snow index (NDSI) and additional thresholds have been accomplished for the two target lakes for October 2016 to April 2017. Considering both the VIIRS I-band reflectance and thermal infrared-derived LSWT were a satisfying approach for lake ice detection, even for small lakes (e.g. Lake Sils). The implemented two-step-approach used these results to determine, without user-interaction, the thresholds to retrieve ice phenology. Finally, this results in an automatic determination of duration of ice cover (phenology) due to the identification of first/last day with ice cover, even for small lakes where only four-six 375m-pixels being not affected by the shoreline. Hence, the developments made for this feasibility study were successful and the pre-operational concept of our modular procedure can be part of an operational service. In-situ measurements and processing were performed by EAWAG. There is no direct method to measure ice coverage in lakes. Time series of lake temperatures are often used to assess freezing and melting conditions. Yet, the 0°C boundary condition for ice cover in lakes is limited to a very thin layer at the surface (0 cm) immediately under the ice and is practically nearly impossible to measure with moorings. In this project, we tested different approaches to monitor the ice cover period by analysing the changes in lake dynamics during ice-free and ice covered period. Namely, we observed that (i) the correlation in the temperature time series of two closely vertically located sensors changed when the external forcing are modified by the ice. Similar results can be observed through (ii) wavelet and Fourier analysis of the temporal evolution of a single temperature logger with a noticeable drop in the energy (e.g. temperature fluctuations) during the ice covered period. Lastly, (iii), we evaluated the potential of using a high frequency pressure sensor to track the ice-on/off periods. Finally, the ice phenology was investigated with two different numerical models. A fully deterministic hydrodynamic model (simstrat.eawag.ch) provided information related to the sensitivity of the ice coverage to the meteorological forcing (e.g. mostly wind and air temperature) and lake characteristics (e.g. bathymetry). This model together with a second hybrid model were used to evaluate the long term phenology of frozen lakes and will allow estimate of the change in phenology under climate change conditions (with CH2018 dataset). The results of ice-on/off dates were based on ground-truth, mainly consisting of visual interpretation of Webcam images, which had some deficiencies and cannot be fully trusted. Furthermore, for the satellite data, clouds on and/or close to the ice-on/off dates led in some cases to large errors in the determination of these dates. The comparison of the different monitoring methods was not based on sufficiently extensive and timely co-incident data. In spite of these difficulties, some very valuable conclusions can still be drawn. Small lakes (< 2 km2) are problematic for satellite images with a ground pixel size of about 250-400m and more. Webcams (though in this case with limited tests) and in-situ (temperature-based) measurements showed the best accuracy. Optical satellite data suffer from clouds, cloud mask errors, reflectance variations and other factors, and their accuracy regarding ice-on/off was worse. The two different processing methods (ETHZ and UniBe) of the same VIIRS data had generally similar performance but differences should be investigated. The main problem for all data (apart in-situ) is the separation between transparent/clear ice (usually thin) and water, particularly in the freeze-up (but also less in the break-up) period, because of very similar reflectance. This problem gets worse with reduced spatial resolution of the images. Even careful visual interpretation of image data cannot reliably distinguish between these two cases. Unfortunately, the critical ice-on/off dates involve water and ice (snow usually comes later on ice-on and melts earlier than ice-off) and thus, accurate identification of transparent ice or water is crucial for the accurate identification of ice-on/off dates. While as known, satellite data is the best operational input for global coverage, but partly also in Switzerland, for Switzerland Webcam and in-situ data are very valuable. Further data, than those used in this project, should include radar satellite data (as highest priority to avoid the cloud problem) and additionally optical high spatial resolution satellite data (especially the free ESA Sentinel-1/2 data), better Webcams (especially with Pan-Tilt-Zoom) and extended in-situ measurements (in combination with local, cheap meteo stations and better Webcams). For in-situ observations and generally, cooperation with Swiss Federal and other agencies and stakeholders would be very beneficial.