We present a statistical attribution framework that distinguishes stationary and nonstationary soil moisture–precipitation coupling, separating direct linear effects from indirect nonlinear pathways. This reveals how anthropogenic forcing and internal variability fundamentally restructure hydrological feedbacks, beyond simple trend attribution.
Calibration and uncertainty quantification for deep learning-based drought detection
We introduce calibration and entropy‑based uncertainty quantification for deep learning‑based drought detection, addressing miscalibration and overconfidence. Our method reduces expected calibration error to 0.31% and improves uncertainty‑based decision‑making accuracy from 72.27% to 76.59% by rejecting uncertain samples. Empirical results across European drought events demonstrate enhanced reliability and trustworthiness for climate decision‑making.
XAIDA4Detection: A Toolbox for the Detection and Characterization of Spatio-Temporal Extreme Events
The XAIDA4Detection toolbox, developed within the XAIDA project, offers open‑source ML models (supervised/unsupervised, deterministic/probabilistic, CNN/RNN‑based) for spatio‑temporal detection and localization of extreme events like tropical cyclones, heatwaves, and droughts. It provides probabilistic heatmaps and is designed for users with basic Python/DL skills. This presentation highlights its adaptability to diverse extreme‑event use cases.
Hybrid Recurrent Neural Network for Drought Monitoring
We propose a hybrid deep learning model for drought detection that integrates climatic indices (multi‑scale SPEI, precipitation, temperature) with data‑driven approaches, using a dual‑branch RNN with convolutional lateral connections. Experiments show it outperforms both standalone indices and purely data‑driven models, demonstrating the potential of hybrid systems that combine domain knowledge with learning.
Attention-based second-order pooling network for hyperspectral image classification
We propose an attention‑based second‑order pooling network (A‑SPN) for hyperspectral image classification, addressing limitations of existing DL methods that ignore higher‑order statistics and suffer from complex hyperparameter tuning. The model uses a first‑order feature operator and an attention‑based second‑order pooling operator, enabling end‑to‑end learning of discriminative features without cumbersome tuning. Experiments on three datasets show A‑SPN outperforms state‑of‑the‑art methods in accuracy, generalization, convergence, and efficiency.