Della Rocca, Fabio (2025) Inference of Water Deficit Indices of Crop and Vegetation from Satellite Hyperspectral Infrared Sensors. [Tesi di dottorato]

Documento PDF Fabio_Della_Rocca_38_COMPLETO.pdf Download (104MB)

Abstract

Global warming is intensifying extreme weather events worldwide, amplifying droughts, floods, and wildfires. In 2024, the global mean temperature reached 1.55 ± 0.13 °C above pre-industrial levels, surpassing the Paris Agreement threshold. Among all climate-related hazards, drought is one of the most complex phenomena and also one of the most difficult to define and monitor. Assessing drought events is crucial and satellite observations can provide significant support thanks to their large spatial coverage and continuous data supply. Vegetation is often among the first systems to exhibit stress because of its high sensitivity to water availability. However, since drought events generally cover vast areas, vegetation monitoring has to be done remotely, traditionally through Vegetation Indices (VIs). VIs describe remotely sensed vegetation properties such as photosynthetic activity, canopy structural variations and water content. In this context, a new Water Deficit Index (WDI) was introduced, defined as the linear difference between the Land Surface Temperature (LST) and the Dew Point Temperature Td. This index gives a compact and physically interpretable measure of surface heating relative to ambient dryness and has already proven effective in detecting water stress in plants. The computation of the index is straightforward, but the retrieval of the required variables is more complex. Specifically, the retrieval of remotely sensed Td is the most challenging. Unlike LST, which is a radiometric “skin” property that can be derived from a few Thermal Infrared (TIR) channels, Td is an atmospheric parameter typically obtained from vertical profiles, thus requiring hyperspectral infrared sensors. Thanks to its high spectral resolution, the Infrared Atmospheric Sounding Interferometer (IASI) is the only sensor capable of simultaneously retrieving LST, Td, and several other atmospheric parameters. IASI combines excellent spectral and temporal resolution (8461 channels and two overpasses per day per platform) with a spatial resolution of approximately 12km at nadir. These characteristics have enabled the development of a retrieval scheme that simultaneously estimates the LST, the full emissivity spectrum, vertical profiles of temperature, water vapour, and ozone mixing ratio, as well as atmospheric trace gases concentrations and, as indirect products, the relative humidity and Td. The complete forward/inverse processing chain for IASI data is implemented in a software package called φ-IASI, a FORTRAN-based inverse radiative transfer code for the optimal estimation of the thermodynamic state of the atmosphere. The retrieval scheme was tested over two Italian regions: the Po Valley region (northern Italy) and Basilicata region (southern Italy). Unfortunately, the retrieval of geophysical parameters from IASI radiances is limited by the nature of the sensor. Since IASI is a TIR sounder, it cannot penetrate thick cloud layers (clouds are opaque in the TIR), resulting in reliable retrievals only for clear-sky pixels. This introduces missing data, requiring an interpolation step to transform the sparsely populated, non-uniform Level 2 (L2) observations into evenly-gridded maps Level 3 (L3). In this work a downscaling–fusion methodology was applied. Several Machine Learning (ML) algorithms with complementary strengths have been explored, including Random Forest (RF), Gradient-Boosted Trees (GBT), Neural Network (NN), and Gaussian Process Regression (GPR), and Stacked Regression. Specifically, the models were developed to predict LST and Td at the locations of the L2 satellite observations, using a set of covariates derived from other satellite sensors, vegetation products, reanalysis datasets, and territorial and geographical information. All input datasets were spatially and temporally co-located with the L2 observations to ensure consistency. Subsequently, the predicted LST and Td fields were downscaled by exploiting the higher spatial resolution of the covariates, producing estimates over a regular 0.05° grid, thereby effectively doubling the spatial resolution of the original data. From the downscaled L3 LST and Td, the WDI was computed, and its robustness was assessed by validating LST and Td individually. Validating LST products is challenging because only a few ground stations exist worldwide. Given the strong spatial heterogeneity of LST, the evaluation focused on the method’s ability to capture finer details and small variations in LST at the target resolution. In this context, the L3 outputs were compared against the LST from the Moderate Resolution Imaging Spectroradiometer (MODIS). Conversely, Td validation was carried out using data from the regional Agenzia Regionale per la Protezione dell’Ambiente (ARPA) network, which provides reliable in situ observations across Italy. Results indicated that the downscaling–fusion methodology significantly improved spatial sampling, while maintaining very low errors. Specifically, the comparison with MODIS for LST and the validation against ARPA stations for Td showed a better agreement for the L3 product than for the L2, confirming that the enhanced spatial resolution effectively captured the underlying spatial variability. The ability of the WDI to capture plant water deficits was further confirmed by comparison with other remotely sensed water stress indicators, namely the Normalized Difference Moisture Index (NDMI) and Surface Soil Moisture (SSM). Finally, the focus shifted towards the projections of the WDI. Understanding how vegetation water stress may evolve under climate change requires more than a snapshot of recent events. It is important to identify long-term trends and determine whether it is becoming a more frequent event. Future climate projections provide a key tool for this analysis. However, since most future projections are based on precipitation-based indices, WDI-based projections provide additional value by offering a complementary perspective. To project the WDI, an ensemble of 13 EURO-CORDEX simulations was used, considering both an intermediate (RCP 4.5) and a worst-case (RCP 8.5) scenario. The analysis focused on the Italian peninsula and was subsequently restricted to the Po Valley and Basilicata regions. The projections indicated consistent positive WDI anomalies across Italy in all seasons and under both scenarios, suggesting an increase in water stress. These anomalies are particularly pronounced in the Po Valley, the islands, and Basilicata, especially under the worst-case scenario.

Downloads

Actions (login required)

Modifica documento