From pixels to charts – using remote sensing for the climate change indicator “lake ice” in alpine lakes

Lake ice is an important feature for many limnic ecosystems. Periodic ice cover influences biological, chemical and physical processes in lakes during the cold climate period. Additionally, ice cover also affects processes outside the ice period, for instance, lake water temperature, timing of spring bloom, primary productivity and mixing regimes. In the face of climate change, many regions experience shifting lake ice cover. The observed and projected loss significantly affects lake ecology but also cultural ecosystem services. In the alpine region of Germany, people associate personal memories, sportive activities and many other aspects with lake ice. Lake ice is connected to society and also strongly affected by climate change. Therefore, it well suits as an indicator to communicate climate change. The region, however, lacks systematic measurements and data on lake ice cover.

In our project, we therefore aimed at developing a remote sensing approach to create a comparable data basis for a climate change indicator on lake ice. Our case study analysed six lakes between 700 and 2000 m AMSL. We generated a data set on lake surface characteristics (water, ice, snow, transparent ice etc.) using public webcam imagery as independent source. The data set was used to train and validate random forest classifiers for Setinel-1 A/B, Sentinel-2 A/B and Landsat 8 imagery. We excluded Sentinel-1 data, which were acquired at wind speeds > 1 m/s (ERA5-LAND). Thus, we prevented erroneous classification of rough waters. The validation revealed very high accuracies with balanced overall accuracies around 0.99, which is misleading. The high accuracies result from how we designed the ground the data since we only used data with labelling under high certainty. In this region, mapping lake ice faces the challenge of multiple freezing in thawing processes within the ice period. We therefore, implemented an air temperature (ERA5-LAND) filter to check the plausibility of classification results.

The final classification results differentiated binary between ice and no-ice pixels. From this data, we defined ice-days with at least 80 % ice cover on a lake. To build the indicator, we divided the monthly sum of ice days by the number of valid image acquisitions. Thus, the indicator also accounts for varyingly available satellite data.

With covering currently seven ice periods, the time series is relatively short for a climate change indicator. The approach may also be transferred to archived imagery whereas lacking ground truth data remain challenging. The small size of (0.2 - 3 km²) complicates the usage of large scale sensors such as MODIS. Thus, combining data from five satellites resolving at 10 – 30 m allowed to generate comparable and spatially explicit data on ice cover of these lakes for the first time.