Active volcanoes present significant hazards at various geographic scales, and volcanic ash clouds produced during explosive eruptions are key factors affecting human health, climate, and aviation safety. Accurately capturing the spatio-temporal evolution of volcanic ash clouds is critical for understanding their dynamics and improving predictive capabilities. Given the highly dynamic nature of explosive eruptions, volcanic clouds can form, expand, and disperse rapidly. Therefore, high-temporal-resolution geostationary satellite data, such as that from MSG-SEVIRI, is indispensable for near-real-time monitoring of these events. Here, we propose VONA-DL (Volcano Observations Nowcasting for Aviation using Deep Learning), a purely data-driven approach to short-term forecast, i.e., nowcast, volcanic gas and ash dispersion emitted by Mt. Etna, by means of a deep learning-based approach to process multispectral Satellite Image Time Sequences (SITS). The model was trained and evaluated using a systematically curated dataset of 91 eruptive events spanning 17 years and classified into four operational scenarios. VONA-DL was able to learn known short-term behaviors without the explicit formulation of physical processes or the specification of atmospheric and eruptive input parameters. In particular, the model reproduced distinct transport, persistence, and removal patterns of ash and SO2 dominated volcanic clouds, providing evidence of its ability to implicitly learn physically meaningful features directly from satellite observations. Once trained, VONA-DL performs inference in approximately 0.005 s per sample, enabling rapid, observation-driven nowcasting as new satellite observations become available. These results demonstrate the potential of VONA-DL as a computationally efficient tool complementary to physics-based dispersion models for short-term volcanic cloud forecasting.
This study presents an integrated multi-satellite assessment of water quality in small high-Andean lakes (SHAL) and a reservoir in the Central Andes. Sentinel-2 MSI and Landsat 8 and 9 OLI imagery were used to evaluate multiple atmospheric correction methods and develop a framework for monitoring remote mountain lakes. Water quality and trophic state were investigated in seven SHAL and one reservoir within the Misicuni and Chocaya watersheds, Bolivia. The study systems are located between 4000 and 4600 m a.s.l. and range in area from 0.01 to 10 km2, whereas the reservoir has an elevation of 3774 m a.s.l. Field samples were combined with multi-satellite observations to estimate chlorophyll-a, transparency, turbidity, and chromophoric dissolved organic matter. Random Forest models were developed using spectral data alone and in combination with landscape and hydroclimatic predictors, while XGBoost was evaluated as a complementary algorithm. Integrating catchment descriptors improved model performance, yielding cross-validated predictions for water-quality variables, while uncertainty estimates were generated for mapped products. Atmospheric correction methods performed differently among water-quality variables, indicating that no single method performs consistently best across variables. Predictions revealed spatial and temporal patterns associated with bofedales (high-Andean wetlands), antecedent precipitation, elevation, and anthropogenic land cover. Remote sensing-derived trophic state and dissolved organic carbon estimates were consistent with field observations, enabling assessment of ecological condition and the spatial distribution of carbon pools across SHAL. The proposed framework provides a scalable approach for monitoring high-Andean lakes, supporting ecosystem-based management, land-use planning, and watershed decision-making where water quality observations remain sparse.
Information about the extent and type of current fuels, informed by vegetation height, can contribute to the management of bushfire-prone areas, particularly when used to support bushfire behaviour modelling. To apply this across a large country like Australia, nationally consistent and up-to-date vegetation height data are required, which are currently lacking. Satellite LiDAR sensors, such as the Global Ecosystem Dynamics Investigation (GEDI) instrument, measure above-ground profiles at spatially discrete points. This can be combined with wall-to-wall Earth observation data to estimate vegetation height across the continent. In this study, GEDI data are combined with Landsat annual products, L-band Synthetic Aperture Radar (PALSAR) annual mosaics, a Digital Elevation Model and climate data through a single machine learning model to predict vegetation height (top-of-canopy) at an annual frequency. The 98th percentile relative height (rh98) from GEDI was selected to represent top of canopy, however since the sensitivity of rh98 reduces for low vegetation cover, which is important when estimating bushfire fuel types, an adjustment was applied to rh98 for heights <3 m using the fractional cover of bare soil. Following the assessment of a selection of machine learning algorithms, an extreme gradient boosting (XGBoost) model was selected to produce Australia-wide vegetation height at a 30 m pixel size for 2019 to 2024. When compared to an airborne LiDAR canopy height model the XGBoost modelled vegetation height had a root mean square error (RMSE) of 5.6 m, mean absolute error (MAE) of 4.1 m and R2 of 0.68. When compared to 397 field plots of vegetation height, the RMSE, MAE and R2 were 4.5 m, 3.1 m and 0.51 respectively. (Although no independent validation data were available along the western side of the continent). Despite these errors, the multi-temporal modelled height was consistent. When compared to two recent global vegetation height products, and an older vegetation height model for Australia, the XGBoost model was found to perform better overall.
