Free-space optical (FSO) downlinks from satellites enable high data rates but are highly sensitive to cloud-induced attenuation and blockage. We present an integrated all-sky imaging framework for optical ground stations that converts station-local sky observations into direction- and lead-time-dependent cloud-free line-of-sight (CFLOS) decision support along predicted satellite links. The framework combines geometric calibration of hemispheric imagery, two-line element (TLE)-based orbit propagation, stereographic remapping into a common processing domain, short-horizon autoregressive sky-frame prediction using a diffusion-based sequence model, cloud/no-cloud segmentation, and a corridor-based CFLOS decision rule along the projected satellite path. The contribution lies in the operational integration of these components into a unified CFLOS-oriented sensing, prediction, and evaluation chain for optical downlink support. The framework is demonstrated at the German Aerospace Center (DLR) optical ground-station site in Trauen and evaluated using a geometry-controlled reference-track protocol across image sequences acquired at 15 s, 30 s, and 45 s cadence. Under this protocol, nowcasting-based CFLOS decisions outperformed a constant-persistence baseline across all evaluated lead times. At a 90 s lead time, the method achieved an F1-score of 0.857 and a balanced accuracy of 0.865, corresponding to gains of +0.083 and +0.089 over persistence, respectively. Positive performance margins are maintained across the full evaluated range up to a 450 s lead time. These results show that all-sky image sequences can be translated into physically interpretable CFLOS decision support and provide a basis for future network-level site-selection and handover strategies.
Deriving land surface temperatures (LST) from aerial thermography requires surface emissivity information, which is typically assumed uniform despite considerable variation across urban materials. We present PROMETHEUS, a workflow that uses a fine-tuned Large Vision Model (LVM) to produce city-scale material classification at airborne resolution. This classification enables emissivity-based LST estimation following the GRAZ method, which uses three-dimensional Monte Carlo sampling to determine view factors for reflected thermal radiation and models elevation-dependent atmospheric transmittance, upwelling and downwelling radiation. We applied this workflow to a 100×100 km area centred on Klagenfurt, Austria, where thermal infrared imagery at 1 m resolution was acquired on August 10-11, 2024 during a summer heat period, with daytime and nighttime flights at 1600 m altitude. A team of 12 surveyors collected concurrent in-situ land and water surface temperatures across 13 stations throughout the city. Using existing 5 cm RGB and near-infrared orthoimagery combined with photogrammetric building segmentation, expert annotators labelled rooftop materials across 30 classes via a collaborative platform with a standardized material guide. These labels were used to fine-tune an LVM that then classified materials across the full study area. The output was merged with municipal land cover data and converted to emissivity values using a look-up table derived from spectral libraries. Atmospheric parameters were obtained from ECMWF profiles. Comparison with in-situ measurements shows improved LST retrieval relative to uniform emissivity assumptions, particularly for low-emissivity surfaces such as metal roofing. This workflow demonstrates a practical approach for scaling limited expert annotations to city-wide material mapping.
Explicit 3D radiative transfer captures the complex spatial variability of global horizontal irradiance (GHI) under shallow cumulus clouds, but is computationally prohibitive for real-time operational use. We address this by introducing a simulation-to-reality framework that emulates 3D radiative transfer via a neural network trained on synthetic data to translate multi-view all-sky imagery into 2D GHI maps. To produce the training data, we ran large eddy simulations at 50 m horizontal resolution for 10 selected cloud dynamic days (April to July 2022) over a 14 km x 14 km domain. The resulting cloud fields were coupled with Monte Carlo ray tracing to render synthetic all-sky images from virtual camera locations matching the Eye2Sky camera network and to compute the corresponding GHI maps. Two datasets are generated, (1) raw synthetic renderings and (2) enhanced renderings with camera-specific characteristics from the real-world. Identical image-to-image neural networks are trained on these datasets and applied to real Eye2Sky imagery, with predicted GHI maps validated against co-located pyranometers. By incorporating sensor-specific characteristics, we quantify the benefit of reducing the simulation-to-reality gap and assess whether synthetic pre-training using neural network emulations can support operational solar irradiance mapping as an alternative to computationally expensive physical simulations.
All-sky imagers (ASIs) used in solar energy meteorology face an extreme intra-image dynamic range, with the circumsolar neighborhood orders of magnitude brighter than the diffuse dome. Many operational ASI pipelines address this gap with high-dynamic-range (HDR) bracketing inside the camera’s image signal processor (ISP), i.e., after demosaicing and color processing in a nonlinear 8-bit RGB domain. Near the Sun, such ISP-domain HDR can down-weight the shortest exposure, retain clipped or near-clipped samples from longer frames, and compress highlight contrast, thereby increasing circumsolar saturation and flattening aureole gradients. A radiance-linear HDR fusion in the sensor/RAW domain (RAW–HDR) is therefore contrasted with the vendor ISP-based HDR mode (ISP–HDR). Solar-based geometric calibration enables Sun-centered analysis. Paired, interleaved acquisitions under clear-sky and broken-cloud conditions are evaluated using two circumsolar performance criteria per RGB channel: (i) saturated-area fraction in concentric rings and (ii) a median-based radial gradient in defined arcs. All quantitative analyses operate on the radiance-linear HDR result; post-merge tone mapping is only used for visualization. Across conditions, ISP–HDR exhibits roughly double the near-saturation within 0– of the Sun and about a three- to fourfold weaker circumsolar radial gradient within 0– relative to RAW–HDR. These findings indicate that radiance-linear fusion in the RAW domain better preserves circumsolar structure than the examined ISP-domain HDR mode and thus provides more suitable input for downstream tasks such as cloud–edge detection, aerosol retrieval, and irradiance estimation.