Predicting Children's Climate Hazard Exposure and Vulnerability: A Reproducible Multi-Task Deep Learning Benchmark Using the UNICEF Children’s Climate Risk Index

Abstract Children bear a disproportionate share of the burden of climate-related hazards, yet the data-driven tools used to quantify where they are most at risk remain dominated by additive index methods and conventional regression. Reliable, jointly-estimated measures of hazard exposure and vulnerability are a prerequisite for targeting adaptation finance and child-sensitive disaster risk reduction. We introduce the ResNet-FT-Transformer with Cross-Attention Hydration (MHA-Dense), a deep tabular architecture that tokenizes continuous socio-environmental indicators into dense embeddings, refines them through four residual transformer blocks combining multi-head self-attention (MHA-Dense) with a gated Cross-Attention Hydration module that injects a learned global context into local feature tokens, and predicts two targets simultaneously through dedicated task heads. Using the country-level UNICEF Children’s Climate Risk Index (CCRI) data (n = 176 countries and territories; nine demographic, hazard, and service-deprivation indicators), the model jointly estimates the Hazard Exposure Score and the Child Vulnerability Score within a single multi-task network. On a held-out test set the framework attains coefficients of determination of R² = 0.727 (MAE = 0.877, RMSE = 1.162) for hazard exposure and R² = 0.818 (MAE = 0.843, RMSE = 1.048) for child vulnerability, performance that is statistically comparable to strong gradient-boosted tree and neural baselines whose bootstrap confidence intervals overlap throughout. A controlled benchmark against six baselines (Extra-Trees, Random Forest, XGBoost, LightGBM, an SVM, and a multilayer perceptron) shows that, on small low-dimensional indicator data, gradient-boosted ensembles retain the strongest point estimates, a result we report transparently and situate within the tabular deep-learning literature. The principal contributions are an interpretable, unified multi-task architecture for child climate-risk modelling, the MHA-Dense and Cross-Attention Hydration mechanisms, and an openly reproducible benchmark with uncertainty quantification to guide model selection in disaster-risk applications.