D'Ercole, Riccardo (2025) Deep Learning models for short-term Drought Prediction in the Horn of Africa. [Tesi di dottorato]
![[thumbnail of Thesis_DErcole_Riccardo.pdf]](https://www.fedoa.unina.it/style/images/fileicons/other.png "Thesis_DErcole_Riccardo.pdf") |
Documento PDF
Thesis_DErcole_Riccardo.pdf Download (10MB) |
| Tipologia del documento: | Tesi di dottorato |
|---|---|
| Lingua: | English |
| Titolo: | Deep Learning models for short-term Drought Prediction in the Horn of Africa |
| Autori: | Autore Email D'Ercole, Riccardo riccardo.dercole@unina.it |
| Data: | 3 Giugno 2025 |
| Numero di pagine: | 89 |
| Istituzione: | Università degli Studi di Napoli Federico II |
| Dipartimento: | Biologia |
| Dottorato: | Intelligenza artificiale Area Agrifood e ambiente |
| Ciclo di dottorato: | 37 |
| Coordinatore del Corso di dottorato: | nome email Loreto, Francesco francesco.loreto@unina.it |
| Tutor: | nome email Sanò, Paolo [non definito] Casella, Daniele [non definito] |
| Data: | 3 Giugno 2025 |
| Numero di pagine: | 89 |
| Parole chiave: | Deep Learning; drought prediction; geostationary satellite; agricultural drought; vegetation |
| Settori scientifico-disciplinari del MIUR: | Area 02 - Scienze fisiche > FIS/06 - Fisica per il sistema terra e il mezzo circumterrestre Area 04 - Scienze della terra > GEO/12 - Oceanografia e fisica dell'atmosfera Area 01 - Scienze matematiche e informatiche > INF/01 - Informatica |
| Informazioni aggiuntive: | Dottorando del 37 ciclo |
| Depositato il: | 27 Ott 2025 15:14 |
| Ultima modifica: | 02 Set 2026 08:08 |
| URI: | https://www.fedoa.unina.it/id/eprint/16763 |
Abstract
This thesis advances the monitoring and forecasting of vegetation dynamics and drought impacts in the Horn of Africa by leveraging high-temporal-resolution satellite data and Machine Learning models. The first study introduces a novel approach to derive daily Normalized Difference Vegetation Index (NDVI) time series from the SEVIRI geostationary sensor, mitigating cloud contamination on vegetation pixels. The resulting dataset exhibits similar performance to other existing vegetation products (e.g., MODIS) in capturing regional climatology and detecting vegetation anomalies with the added benefit of the increased temporal resolution. Building on this foundation, the second study compares the derived high-frequency NDVI dataset with gloabal and regional precipitation products in detecting meteorological and agricultural droughts across Ethiopia, Somalia, Kenya, and Djibouti. The analysis highlights the value of daily temporal resolution in capturing short-term fluctuations and improving drought characterization across different land cover and soil types. Moreover, the study emphasizes a trade-off between accuracy and precise detection of meteorological drought events for precipitation indices at different accumulations. The final study integrates the vegetation dataset with meteorological data from the reanalysis model ERA5 and soil moisture from ASCAT radiometer to develop a series of Deep Learning models for short-term drought forecasting. A diffusion-based probabilistic model, conditioned on prior vegetation and meteorological history, demonstrates superior performance over traditional recurrent architectures, particularly under extreme drought scenarios. Collectively, these studies provide an end-to-end pipeline from data reconstruction to predictive modeling that enhances early warning capabilities in climate-vulnerable regions.
Downloads
Downloads per month over past year
Actions (login required)
 |
Modifica documento |
