L’exploitation de l’énergie solaire soulève des défis liés à la nature variable de la res- source concernée : le rayonnement solaire. Son comportement intermittent est un problème pour la gestion des centrales photovoltaïques et du réseau électrique. L'une des solutions largement envisagées est la prévision de la production photovoltaïque à différents horizons.L'objectif de cette thèse est d'explorer de nouvelles voies pour améliorer les prévisions existantes du rayonnement solaire, pour des horizons allant de quelques minutes à quelques heures, en exploitant les synergies possibles entre les mesures pyranométriques, les images hémisphériques du ciel prises depuis le sol et les images acquises par les satellites météorologiques géostationnaires. Ces deux types d’images ont des couvertures spatiales, des résolutions spatio-temporelles et des points de vue très différents.L’approche proposée dans cette thèse exploite cette différence de points de vue afin d’affiner la géolocalisation des nuages en 3D par stéréoscopie, dont l’évolution des ombres peut alors être estimée et prévue. Un simulateur géométrique de la méthode a été développé et permet d’en identifier certains avantages et limitations. La géolocalisation des nuages appliquée à des données réelles a permis d’élaborer des schémas d’estimations et de prévisions prometteurs du rayonnement solaire incident. Enfin, pour compléter l’analyse usuelle de ces performances de prévision, deux nouvelles métriques ont été proposées dans l’optique de quantifier deux notions essentielles : le respect du suivi des rampes et l’alignement temporel de la prévision par rapport à la mesure. / The exploitation of solar energy raises challenges related to the variable nature of the resources involved: the incident solar irradiance. Its intermittent behavior is an is- sue for photovoltaic power plants and grid management. One of the solutions that have been widely considered is the forecast of photovoltaic production at different time horizon.The aim of this thesis is to explore new ways for improving the existing solar irradiance forecasts, for horizons ranging from the present moment to few hours, by exploiting possible synergies between pyranometric measurements, hemispherical images of the sky taken from the ground and images acquired by geostationary meteorological satellites. These two types of images have completely different spatial coverage, spatio-temporal resolutions and are taken from two different locations.The proposed approach in this thesis exploits this difference in points of view in order to geolocate the clouds in 3D by stereoscopy, which shadows’ location and motion can then be estimated and forecasted. A geometric simulator of the method has been developed to identify some of the advantages and limitations of this approach. The geolocation of clouds applied to real data made it possible to develop promising estimates and forecasts of incident solar irradiance. Finally, to complete the usual analysis of forecasting performances, two new metrics have been proposed in order to quantify two essential notions: the ability to monitor the ramps and the temporal alignment of the forecast with the measurements.

Snow interception is one of the most important process of the hydrological balance of river basins. Measuring of snow interception is a very complex activity, therefore, models are frequently used to calculated snow interception from the vegetation structure and measured meteorological variables. A field research has been carried out in the Ptačí Brook basin in the Sumava Mts. to describe the canopy structure of the spruce forest using hemispherical images taken in the winter season 2020/21. The vegetation characteristics are essential for modelling of the snow interception. The mean Leaf area index calculated from the hemispherical images at the study plot reached 2.34 with the respective canopy closure equal to 86.16%. These values were further used as input values for the calculation of seasonal cumulative snow interception at the study plot for the winter season 2020/21. The original, more complex model was compared with two, less complex equations. Consequently, the model was applied to four consecutive winter seasons 2017-2021. The efficiency of the snow interception (a proportion of the intercepted snow to total snowfall water equivalent) ranged from 36.85% to 45.81% depending on the study season. The snow interception efficiency was considerably higher in the last winter season compared to...