2026

Acknowledgement
July 07, 2026 Dissertation Research Hybrid Probabilistic Deep Learning for Drought Monitoring
Huge thanks to Prof. Gustau Camps-Valls and Prof. Miguel Ángel Fernández-Torres for their mentorship throughout this work. You can find the defense slides here!
Long-term Soil Moisture Forecasting: Land–atmosphere Process → Physical consistency
July 01, 2026 Drought Forecasting Soil Moisture Deep Learning Physical Regularization
Soil moisture (SM) droughts pose significant threats to water resources, agricultural productivity, and food security. To mitigate these impacts, reliable long-term drought forecasts are essential.

2025

Reliable probabilistic prediction: Calibration → Uncertainty → Reliability
July 01, 2025 Drought Prediction Deep Learning Uncertainty Quantification Calibration
Traditional drought models, such as process-based and drought indices, often carry bias and uncertainties stemming from mismatches between observational data and process assumptions, as well as from climate variability.
EU Horizon XAIDA — Advanced AI for Detecting and Understanding Extreme Events
July 01, 2025 Toolbox Extreme Events Deep Learning Explainable AI
We developed The AIDE (open-source toolbox) for Detecting and Understanding Extreme Events!

2023

Extreme Event Detection: Hybrid AI → Generalization
July 01, 2023 Drought Monitoring Deep Learning Domain Knowledge
Recognizing real-world low-likelihood high-impact drought events is exceptionally significant. Historically, the impacts of extreme droughts have spanned across ages, triggering devastating consequences such as the collapse of ancient civilizations, forced human migrations, and massive loss of life.

2021

Optimizing Hyperspectral Image Classification: Exploring GCNs and Attention Mechanisms
June 01, 2021 Hyperspectral Imaging Deep Learning Graph Convolutional Networks Attention Mechanisms Few-Shot Learning
Deep learning (DL) has profoundly transformed Hyperspectral Image (HSI) classification by capturing robust features in an end-to-end manner. However, traditional DL models face significant hurdles, including the need for massive amounts of labeled data and the manual tuning of complex structural hyperparameters.