Current groundwater prediction models often exhibit low accuracy and complex parameter adjustment. To tackle these limitations, a novel prediction model, called improved Aquila optimizer bi-directional long-term and short-term memory (IAO-BiLSTM) network, is proposed. IAO-BiLSTM optimizes the hyperparameters of the BiLSTM network using an IAO algorithm. IAO incorporates three novel enhancements, including population initialization, population updating, and global best individual updating, to overcome the drawbacks of current optimization algorithms. Before making predictions, the challenge posed by the highly nonlinear and non-stationary characteristics of groundwater level signals was addressed through the application of a wavelet multi-scale analysis method. Using a landslide site in Zhejiang Province as an example, a monitoring system is established, and continuous wavelet transform, cross-wavelet transform, and wavelet coherence analysis are employed to perform multi-scale feature analysis on a 2-year dataset of rainfall and groundwater depth. The findings reveal that the groundwater depth of monitoring holes exhibits similar high energy resonating periods and phase relationships, strongly correlating with rainfall. Subsequently, IAO-BiLSTM is employed to predict groundwater depth, and its results are compared with seven popular machine learning regression models. The results demonstrate that IAO-BiLSTM achieves the highest accuracy, as evidenced by its root mean squared error of 0.25.

Debris flows represent a persistent challenge for disaster prediction in mountainous regions due to their highly nonlinear and multivariate triggering mechanisms. This study proposes an explainable deep learning framework, the Improved Kepler Optimization Algorithm-Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention (IKOA-CNN-BiGRU-Attention) model, for precise debris flow hazard prediction in the Yunnan section of the Nujiang River Basin, China. The model is developed and validated using data from 159 debris flow-prone gullies, integrating deep convolutional, recurrent, and attention-based architectures, with hyperparameters autonomously optimized by IKOA. Model explainability is enhanced using SHapley Additive exPlanations (SHAP), which quantify the influence of key factors. The IKOA-CNN-BiGRU-Attention framework consistently outperforms 13 benchmark models, achieving a root mean square error of 2.33 × 10⁻⁶, mean absolute error of 1.51 × 10⁻⁶, and mean absolute percentage error of 0.006%. The model maintains high stability across 50 repeated experiments, strong resilience to 20% input noise, and robust generalizability under five-fold cross-validation. Interpretability analysis identifies potential source energy and maximum 24-hour rainfall as primary determinants and uncovers a dual-threshold physical mechanism underlying debris flow initiation. These findings provide a quantitative basis for adaptive early warning and targeted risk mitigation, and establish a transferable framework for explainable geohazard prediction.