Environmental Sensing and Borrower Feature Interaction for Uncertainty-Aware Loan Risk Prediction Under Extreme...

Extreme climatic events, including severe heatwaves, prolonged droughts, and anomalous flooding, can substantially disrupt agricultural production stability and further increase the risk of agricultural loan default. To address the limitations of conventional credit assessment in fully exploiting environmental sensing information, the scarcity of extreme-disaster samples, and fluctuations in sensing quality, a robust default risk prediction framework integrating agricultural environmental sensing observations, borrower-feature interactions, and uncertainty-aware learning is proposed. First, environmental exposure features and borrower-specific attributes are separately extracted through a dual-branch encoding network, and bidirectional cross-attention is employed to capture nonlinear interactions between climatic conditions and borrower risk. Subsequently, a climate-constrained disaster scenario augmentation mechanism is introduced to expand extreme-environment samples under regional, seasonal, crop-growth-stage, and inter-variable relationship constraints. A stabilized weighting strategy is further constructed by combining stochastic prediction with uncertainty estimation, thereby reducing the influence of highly noisy samples and anomalous observations on model training. Experimental results demonstrate that the proposed model achieves a PR-AUC of 0.791, a ROC-AUC of 0.909, and an F1-score of 0.781 under the combined evaluation of normal and extreme climatic conditions, while attaining a Brier Score of 0.098 and an ECE of 0.027. Under degraded sensing conditions involving missingness, noise, and bias, the model maintains high discriminative capability (PR-AUC of 0.758, ROC-AUC of 0.891, and F1-score of 0.756) and well-calibrated risk estimates (Brier Score of 0.112 and ECE of 0.031). Disaggregated evaluations across perturbation types, sensitivity analyses regarding spatial matching and synthetic labeling, external temporal holdout and cross-regional validations, and post-hoc feature attribution analyses further confirm the robustness and practical interpretability of the framework. Overall, the proposed framework consistently outperforms Logistic Regression, SVM, Random Forest, XGBoost, LightGBM, DNN, LSTM, Deep Ensemble, and the previous model.