ICE DETECTION IN SWISS LAKES USING MODIS DATA

In this research, we process low spatial resolution satellite images (MODIS) for integrated multitemporal monitoring of ice in selected lakes in Switzerland. Lake ice is important for climate research and is considered one of the Global Climate Observing System’s (GCOS) Essential Climate Variables (ECVs). The aim of our project is to detect whether a lake is frozen or not. Four of the target lakes are Sihl, Sils, Silvaplana and St. Moritz, showing different characteristics regarding area, altitude, surrounding topography and freezing frequency, describing cases of medium to high difficulty. From the satellite sensor MODIS with daily temporal resolution, several spectral channels are used, both reflective and emissive. The low-resolution MODIS bands with 500m and 1000m Ground Sampling Distance (GSD) are super-resolved to 250m resolution and co-registered prior to the analysis. Digitized lake outlines after generalization using Douglas Peucker Algorithm are back-projected on to the image space. As a pre-processing step, the absolute geolocation accuracy of the lake outlines is corrected by matching the projected outlines to the images. Only the cloud-free pixels which lie completely inside the lake (clean pixels) are analyzed. We formulate the lake ice detection as a two-class (frozen, non-frozen) semantic segmentation problem, but also analyze the three-class distinction where bare ice is separated from snow-covered ice, due to their different spectral properties. The most useful MODIS channels to solve the problem are identified with xgboost, while the classification is done with (non-linear) Support Vector Machines (SVM). The proposed method is tested on MODIS data from the cold winter 2011-12 and summer 2012 and we achieve >95% accuracy on all the four target lakes.