Natural disaster impact assessment is of the utmost significance for post-disaster recovery, environmental protection, and hazard mitigation plans. With their recent usage in landslide susceptibility mapping, deep learning (DL) architectures have proven their efficiency in many scientific studies. However, some restrictions, including insufficient model variance and limited generalization capabilities, have been reported in the literature. To overcome these restrictions, ensembling DL models has often been preferred as a practical solution. In this study, an ensemble DL architecture, based on shared blocks, was proposed to improve the prediction capability of individual DL models. For this purpose, three DL models, namely Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), together with their ensemble form (CNN–RNN–LSTM) were utilized to model landslide susceptibility in Trabzon province, Turkey. The proposed DL architecture produced the highest modeling performance of 0.93, followed by CNN (0.92), RNN (0.91), and LSTM (0.86). Findings proved that the proposed model excelled the performance of the DL models by up to 7% in terms of overall accuracy, which was also confirmed by the Wilcoxon signed-rank test. The area under curve analysis also showed a significant improvement (~4%) in susceptibility map accuracy by the proposed strategy.

Meteorological drought, exacerbated by global climate change, is a recurrent and complex catastrophe that disrupts water resources, agriculture, and economies worldwide. Forecasting future drought conditions is therefore of utmost importance in facilitating effective intervention and adaptation strategies. Artificial intelligence and machine learning techniques have gained attention owing to their potential in forecasting drought severity. However, they are often criticized for their black-box nature and lack of explainability and interpretability. To address this issue, this study explores the potential of local and global explainable artificial intelligence (XAI) tools, including SHAP and LIME, to extract intrinsic characteristics from different drought indices, including Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) and climatic parameters (temperature, relative humidity, and wind speed). These techniques were applied to understand the driving forces behind drought events under different scenarios. SHAP was used to provide global model explanations, revealing how climatic parameters varied in importance across stations and scenarios, while LIME was employed to offer local explanations for specific predictions, highlighting the role of lag time and the significance of each parameter in individual forecasting scenarios. Using data from four meteorological stations in the Konya province of Türkiye, spanning 1965 to 2023, six ensemble-based machine learning models were employed to forecast SPI and SPEI values at 6-month and 12-month scales, resulting in a total of 768 scenarios. Widely used evaluation metrics, namely MSE, RMSE, MAE, MAPE and $$\:{R}^{2},$$ were employed to compare model performances. The prediction accuracy ranged from 0.513 to 0.864 $$\:{R}^{2}$$ values, and CatBoost emerged as the best-performing method in the scenario in which SPEI-12 (t), SPEI-12 (t-1), temperature (T), and wind speed (W) were used as inputs. The global explanation using the SHAP revealed that climatic parameters had varying degrees of impact across stations and scenarios, whereas the influence of lag time was significant in the predictive models. Alongside these considerations, LIME and SHAP methods were applied to investigate the status of the parameters at local and global levels. LIME and SHAP analyses also revealed that T and W parameters played less decisive roles at all stations than SPEI(t) and SPEI(t-1) parameters. The conclusion drawn from the current investigation may offer more explainability when working with machine learning algorithms in drought prediction studies, which could be critical for predicting future drought events.

Landslide susceptibility prediction practices have been increasingly reliant on non-geographically-oriented (i.e., aspatial) machine learning algorithms. While these approaches have exhibited increasing success, they have often faced criticism for their limited consideration of spatial autocorrelations and local variations across geographical space, thereby neglecting the concept of spatial non-stationarity. To fulfill the research gap, this work applies a geographical random forest (GRF) approach, contrasting it with the conventional random forest (RF) algorithm. To this end, the study area, encompassing the Lake Sapanca Basin and its surroundings, was subdivided into 4,452 slope-based mapping units. The effectiveness of both predictive models was then measured by using overall accuracy (OA) and area under the curve (AUC). The results revealed that the GRF (OA = 80.82% and AUC = 85.22%) outperformed the RF algorithm (OA = 75.34% and AUC = 82.50%) by approximately 5% in OA, and demonstrated a 3% improvement in AUC score. The Wilcoxon signed-rank test confirmed significant differences (95% level) between the predictions of both models. The slope parameter emerged as the globally most influential factor, but local interpretations disclosed notable variations in the importance of causative factors contingent upon location. For instance, the curvature parameter was the most important geospatial covariate in around one-third (34.23%) of the slope units, mostly concentrated in the northernmost zones of the study area. On the other hand, elevation was the most important factor for 14.67% of the slope units primarily located in the southern region.