Proposed topics for Masters thesis
(Scroll to the bottom of the page to check out some examples of past Masters theses)
Zonation of rainfall in Switzerland based on variograms
Rainfall in Switzerland varies spatially: the Plateau, the Jura, the Alpine crest, the inner-Alpine dry valleys and the contrast between convective and stratiform rain influence its spatial correlation structure. Existing regionalisations, such as the MeteoSwiss climate regions or the principal-component regions of Baeriswyl and Rebetez (1997), rest on means and seasonal cycles. No zonation exists derived from variogram properties. Operational products either fit one variogram for the whole country (Schiemann et al. 2011) or impose catchment-based sub-regions (Frei and Isotta 2019).
This topic aims at producing the first map of local variogram properties of rainfall in Switzerland and derive from it a zonation of rainfall regimes. It will be based on the CombiPrecip dataset, a radar-gauge combined gridded product at 1 km resolution, available from 2005. The process will involve processing large amounts of data with Python scripts, in collaboration with the cool PhD students in the GAIA Lab. The data are already available on UNIL servers.
References
- Baeriswyl, P.-A. and Rebetez, M. (1997). Regionalization of precipitation in Switzerland by means of principal component analysis. Theoretical and Applied Climatology 58, 31–41.
- Frei, C. and Isotta, F. A. (2019). Ensemble spatial precipitation analysis from rain gauge data. Journal of Geophysical Research: Atmospheres 124, 5757–5778.
- Germann, U. and Joss, J. (2001). Variograms of radar reflectivity to describe the spatial continuity of Alpine precipitation. Journal of Applied Meteorology 40, 1042–1059.
- Schiemann, R. et al. (2011). Geostatistical radar-raingauge combination with nonparametric correlograms. Hydrology and Earth System Sciences 15, 1515–1536.
Automatic detection and geostatistical gap-filling of rain gauge data in the Vallon de Nant
The Vallon de Nant, an alpine catchment in the Vaud Alps, is equipped with three meteorological stations (Auberge, Chalet and Chaux), each with a tipping-bucket rain gauge. Stations data are periodically affected by disruptions (clogging of the funnel, poor signal, etc), and then they record zero or strongly underestimated rainfall. The corrupted periods currently have to be identified by hand, by comparing the station with its two neighbours, before being replaced by missing values. With three stations only a few kilometres apart, these records are a valuable dataset for studying precipitation in a small alpine valley, and losing part of them is a pity.

The goal of this project is to develop and test a workflow that makes this process automatic and that reconstructs the lost data. It has two components:
- Automatic detection of invalid periods. Design a set of statistical tests comparing the three time series (e.g. systematic deviation from the other stations, minimum duration of the anomaly, anomaly ending at a maintenance visit) to flag the periods during which a gauge is clogged. The same tests could later run in near real time on the stations that transmit their data to the lab server, to trigger an alert when a gauge appears to be blocked.
- Gap-filling with multiple-point statistics. Reconstruct the flagged and missing periods using the Direct Sampling (DS) algorithm, a multiple-point geostatistical method that simulates missing values by resampling patterns from the historical record, with the neighbouring stations as auxiliary variables. This approach has been successfully applied to daily streamflow gap-filling in the Volta basin (Dembélé et al., 2019, Journal of Hydrology) and will be adapted here to high-resolution rainfall. The performance of the method will be assessed on synthetic gaps, and the resulting stochastic reconstructions will quantify the uncertainty of the infilled values.
The dataset is cleaned and ready for use from 2018, but some preprocessing is need for earlier data, dating back as early as 2008. The student should be comfortable with programming (typically Python) and interested in data analysis and hydrometeorology; some fieldwork in the Vallon de Nant is possible.
Reference: Dembélé, M., Oriani, F., Tumbulto, J., Mariethoz, G., Schaefli, B. (2019). Gap-filling of daily streamflow time series using Direct Sampling in various hydroclimatic settings. Journal of Hydrology, 569, 573–586.
Improving Random Forests using Kriging‑Based Sample Weights
Random forests are widely used for spatial prediction because they handle nonlinear relationships and complex interactions, but they typically ignore spatial autocorrelation and treat all training samples as equally reliable. In contrast, kriging explicitly models spatial dependence and provides location‑specific interpolation weights and uncertainty estimates (kriging variance). Combining these two perspectives could improve predictive performance and uncertainty quantification, especially in settings where sampling is irregular, clustered, or subject to measurement noise.
The main objective of the project is to develop and test methods that use kriging outputs—particularly kriging weights and variances—to define location‑dependent weights for samples or predictions in random forest models. The student will explore several integration strategies. One strategy is to use kriging weights or kriging variance as sample weights during random forest training, so that samples in areas of high uncertainty or low support exert less influence on the learned model, while samples in well‑supported regions are emphasized. Another strategy is to build a hybrid model in which kriging and random forest each produce a prediction at a location, and these predictions are combined using weights derived from their respective uncertainties, for example via inverse‑variance weighting.
The project will start based on synthetic cases where the spatial distribution of the labels is controlled, before moving to real datasets based on satellite remote sensing. It will also include sensitivity analyses to understand how the proposed methods behave under different conditions, such as varying sampling density and spatial clustering, the presence of measurement noise or label uncertainty.
Classification spatio-temporelle des habitats dans la réserve de la Grande-Cariçaie
La réserve de la Grande Cariçaie (https://grande-caricaie.ch/fr/accueil/), sur les rives du lac de Neuchâtel, abrite une grande diversité d’habitats, et est un lieu d’étude idéal des changements environnementaux. Ce travail consistera à cartographier ces habitats à l’aide de différentes images satellites et aériennes, ainsi que de données de terrain acquises en collaboration avec l’association de la Grande Cariçaie.
Illustration de données d’entrainement utilisées lors de tests préliminaires.

Investigating foundation models for generating synthetic satellite images
There is a scientific interest in generating synthetic satellite images, for instance to complete image time series in case of sensor malfunction, or to generate old images for epochs when satellites were not yet available. Recently, foundation models such as Prithvi-EO-2.0 (https://research.ibm.com/blog/prithvi2-geospatial) have shown potential in solving many Earth Observation problems. This made it possible to solve a range of problems related to the use of satellite data. This project will consist in testing Prithvi-EO-2.0 for synthetic satellite image generation, using a number of test cases for which existing results are available for comparison.
Machine learning-based hydrograph seperation
Streamflow time series of mountainous catchments are a mix of rainfall and snowfall output. It is difficult to unmix them, to estimate the partition between rainfall and snowfall just from looking at the discharge. In this project, you will give it a try with a machine learning algorithm. The alrogithm will be trained in an artificial setting, where we know both the inputs (mixed rainfall/snowfall discharge) and the outputs (unmixed rainfall+snowfall partitions). Once it works on the artifical setting, you can try it out on real discharge time series.

Comparing observed and simulated snow line altitudes
The snow line altitude is crucial for water resources, ecology, and winter tourism, but predicting it accurately is challenging. While remote sensing and webcam imagery provide reliable observations, meteorological forecasts and reanalysis products often struggle to estimate the zero-degree line, a key predictor of the snow line altitude. Local factors such as humidity, temperature stratification, and surface conditions further complicate the prediction of whether precipitation falls as snow or rain.
This project will compare snow line altitudes estimated by meteorological products with observations from webcams and remote sensing imagery. Given the limitations in resolution and temperature accuracy in mountainous terrain, identifying systematic biases in these products could help apply corrections, ultimately improving hydrological and ecological models in these regions.
Cartographie de la granulométrie des éboulis en haute montagne
Dans les régions de haute altitude, de grandes zones sont couvertes d’éboulis. La taille des grains qui composent ces éboulis est un facteur important contrôlant l’occurrence de permafrost et aussi l’apparence de végétation. Une méthode a été mise au point pour cartographier et identifier les grains à partir d’images prises par drones . Ce projet consiste à appliquer cette méthodologie sur une zone étendue des Alpes Vaudoises.
Analyse de séries de débits dans le Val d’Heremence
Ce sujet consistera à analyser de longues séries temporelles de début dans la région du barrage de la Grande Dixence. A partir de données historiques de débit, il est possible de déduire la durée et l’intensité des périodes de fonte de neige, et partant d’estimer les conditions d’enneigement à des époques lointaines dans le passé (i.e. lorsque les données satellite n’étaient pas encore disponibles).
Reconstruction de données d’évaporation synthétiques pour le Bassin de la volta
Dans la région du bassin de la Volta (Afrique de l’Ouest), des millions de personnes dépendent de ressources en eau dont le régime est peu informé en données. Ce sujet consistera à mettre en oeivre des méthodes statistiques développées dans notre groupe pour générer des données d’évapotranspiration synthétiques mais plausibles, pour des périodes où les données satellite de sont pas disponibles (p.ex. dans le passé). Le but final est de pouvoir reconstruire l’état du bassin-versant sur une longue période à l’aide de modèles qui ont besoin de ce genre de données.
Downscaling of Digital Elevation Models with landscape evolution models
High-resolution Digital Elevation Models (DEMs) are often not available for studies of the earth surface. This project aims at testing an approach for downscaling DEMs using the laws governing erosion of a landscape (https://www.sciencedirect.com/science/article/pii/S0169555X12004618).
Etude de la relation entre pluviométrie et altitude à plusieurs échelles
La pluie est typiquement considérée comme corrélée à l’altitude, et cette relation est utilisée de façon générale pour interpoler des données ou pour formuler des modèles. Cependant, des données très locales semblent indiquer que cette relation n’est pas toujours valide lorsqu’on se focalise sur l’échelle de petits bassins-versants et sur des moyennes mensuelles ou saisonnières plutôt que annuelles. Il est possible que dans ces cas, d’autres facteurs peuvent être prépondérants pour expliquer l’intensité des pluies. Ce projet vise à établir de telles relations (ou absence de relations) en analysant des données de stations météorologiques dans les alpes.
Quantify extreme temperature values on the earth surface using Landsat-8 TIRS
Landsat-8 has the highest-resolution long-term thermal sensor. It can allow to spot very local temperature extremes that are undetectable with lower resolution sensors. This project is to detect such extreme surface temperatures and possibly explain them.
Colorization and downscaling of drone imagery
Images acquired by drones often do not have many spectral bands (typically RGB). Multispectral camera can be mounted on drones, but then offer a lower resolution. This topic will apply a pansharpening method to add bands to drone images acquired with a high-resolution RGB camera, based on a lower resolution multispectral acquisition.
It will be based on an approach to statistically reconstruct the missing bands in images, described in this paper:
Gravey M, Rasera L, Mariethoz G (2019). Analogue-based colorization of remote sensing images using textural information, ISPRS Journal of Photogrammetry and Remote Sensing, 147: 242-254, doi: 10.1016/j.isprsjprs.2018.11.003.
Quantifying patterns of change in snow cover
Snow cover in alpine regions is complex. It is driven by many factors, such as altitude, slope, solar exposition, vegetation, among others. Daily satellite images allow observing the snowy areas and their evolution in great details. This project will use such data and morphological operators (https://en.wikipedia.org/wiki/Mathematical_morphology) to determine the processes through which snow appears and disappears throughout a season. Such methods allow defining the countours of the snow areas or their “patchiness”.
Merging radar and gauge measurements using multiple-point geostatistics
There are two main sources of data for operational measurement of rainfall: weather radars and rain gauges. While radar data are very useful to obtain spatial maps and to measure rain in inaccessible areas, it generally produces data that have a relatively low spatial resolution and that present numerous biases. Rain gauges data are generally much more accurate, but only available as point measurements, which need to be interpolated, thus resulting in significant uncertainty. Several methods exist to merge these two sources of data, however they often involve assumptions that result in a misrepresentation of rainfall heterogeneity. Moreover, the final product often does not match what is measured in the rain gauges (e.g. the radar indicates rain where rain gauges measured no rain).
This project will investigate the use of multiple-point geostatistics methods, to generate rain fields that are consistent with the heterogeneity of the radar data, and that are in agreement with the rain gauges. It will involve using existing software and available MeteoSuisse data products.
Investigating the influence of shadows on reflectance spectra
When using remote sensing in mountain areas, shadows can pose significant problems. For a given surface type, the reflectance spectrum can be very different depending on the shadow/sunlit condition. In shadow areas, most of the incident light comes from reflection on surrounding slopes, as well as from skylight (atmospheric effects). As a result, classification of corresponding images can be challenging and often one has to define separate classes for shadown and sunlit areas.
This project will investigate ways to relate the reflectance spectra of shadowy and sunlit areas, for a given surface type. Snow surfaces are of particular interest because the high reflectance implies that surrounding slopes will have a strong influence on shadow areas. The work will involve a comprehensive literature review, followed by a systematic comparison of spectra acquired in mountain regions, for images coming from different satellite platforms, including Sentinel-2, Landsat and MODIS.
Some past masters theses
Candice Dubath (2023) – PDF
Jason Bula (2019) – PDF
Dionys Lugon-Moulin (2019) – PDF
Myriam Eggerschwiler (2018) – PDF
Rachel Bungener (2018) – PDF
Danick Perrin (2018) – PDF
Raphael Vallat (2018) – PDF
Romain Udry (2017) – PDF
