Proceedings Article | 20 November 2024
KEYWORDS: Vegetation, Education and training, Interpolation, Machine learning, Spatial resolution, Data modeling, Soil moisture, Air temperature, Climate change
2023 has replaced 2016 as the warmest year on record since 1850, bringing us closer to the 1.5°C limit set by the Paris Agreement. High temperatures increase the likelihood of extreme events, with heatwaves and drought being prominent among them. Climate change has led to a rise in the frequency of droughts, affecting countries that never experienced them. Assessing drought events is crucial and satellite data can provide significant assistance due to its large spatial coverage and continuous data supply. Based on the Infrared Atmospheric Sounder Interferometer (IASI), we designed a new Water Deficit Index (wdi) that we have already proven useful in detecting drought events. Unfortunately, infrared sensors such as IASI cannot penetrate thick cloud layers, so observations are blinded to surface emissions under cloudiness bringing sparse and not homogeneous distributed data over a given spatial region. To reconstruct a model of the field of interest for the entire surface on a regular grid mesh, interpolation techniques, and spatial statistics to deal with huge data sets are mandatory. In this paper, we exploited the capability of two machine learning algorithms, i.e. gradient boosting and random forest, in converting IASI L2 scattered data to a regular L3 grid. Specifically, we trained a model that can predict the wdi index over a 0.05° regular grid, using data from other sensors as a proxy, including vegetation indices (FVC, NDVI), soil moisture content, and geographic information (digital elevation models, land cover types) as covariates. We applied the methodology over the Po Valley region, which experienced an intense drought in the last three years causing high vegetation and soil water stress. Overall, the methods achieved a significant improvement in spatial resolution, with mean absolute error (MAE) and root mean square error (RMSE) values of 1.28°C and 1.68°C, respectively, enabling efficient regular grid conversion and downscaling.
