| 07 Apr 2026
Applied Machine Learning for Flood Susceptibility Mapping in the Loukkos Basin (Northern Morocco): Validation Using the February 2026 Ksar El Kebir Flood Event
Abstract. Flooding is a recurrent hazard in northern Morocco, where low-lying alluvial plains, strong river–floodplain connectivity, and winter storm sequences combine to produce damaging floods. This study develops an event-informed flood susceptibility assessment for the Loukkos Basin (Ksar El Kebir–Larache floodplain) by integrating satellite-derived flood evidence with geomorphometric and hydro-climatic predictors in a machine-learning framework. A binary flood inventory was derived from Sentinel-1 SAR change detection by contrasting pre-flood (October–November 2022) and flood-phase acquisitions (December 2022), and was validated using Sentinel-2 optical observations and field checks. Nine conditioning factors (elevation, slope, aspect, curvature, distance to rivers, drainage density, TWI, TPI, and CHIRPS-based rainfall) were compiled and standardized on a 10 m grid. Model training used a balanced sample of 220 points (110 flooded/110 non flooded) and was evaluated with a repeated hold-out strategy (10 iterations; 80 % training/20 % testing) using accuracy, precision, recall, and F1-score. Both Random Forest (RF) and Multilayer Perceptron (MLP) produced coherent susceptibility patterns, with the highest classes concentrated along the Loukkos river corridor and downstream floodplains. Mean test performance indicates strong generalization, with MLP outperforming RF (accuracy ≈ 0.909 vs. 0.864; F1 ≈ 0.909 vs. 0.870). Jackknife sensitivity analysis identifies elevation as the leading control for both models, while RF emphasizes terrain metrics (slope, drainage density) and MLP assigns a more balanced importance to topography and hydrological drivers (TPI, rainfall, distance to channel). Notably, flooded areas observed during the January–February 2026 flood episode (Sentinel imagery dated 14 February 2026) largely coincide with zones mapped as high susceptibility, providing an independent, qualitative post-study consistency check. The resulting maps offer a practical inspection tool to support land-use planning and prioritization of mitigation actions across the most flood-prone sectors of the Loukkos Basin.
Transformer, more than meets the eye: A deep learning approach to integrate rainfall time-series in multi-type landslide probability modeling
The paper models landslide type‑specific “pseudo‑probabilities” for the May 2023 Emilia‑Romagna events using slope units (SUs). Rainfall is interpolated to SU centroids and either aggregated over fixed windows or fed as daily/hourly time series into a Transformer; static terrain/geology features go through a DNN. Authors report AUCs >0.90 and state that daily time series generally perform best, with SHAP‑based “Expected Gradients” used for explanation. There are many issues such as followings,
Very Fast Kolmogorov-Arnold Network for Landslide Susceptibility Prediction: A Case Study in Gümüşhane, Türkiye
The paper compares six algorithms (XGBoost, LightGBM, NGBoost, CatBoost, MLP, FastKAN) to map landslide susceptibility in Gümüşhane Province using 20 conditioning factors. After a “multicollinearity analysis,” the authors train models on pixel samples (1,446,696 landslide pixels + the same number of non‑landslide pixels), split randomly 70/30, tune hyperparameters with grid search, evaluate using Accuracy/Precision/Recall/F1, interpret with SHAP, and “validate” by overlaying two 2025 landslides on the FastKAN map. They report 96.54% accuracy for FastKAN and declare it the first KAN application to landslide susceptibility. There are many issues such as followings,
Estimating Temporal Probability of Rainfall-Induced Landslides Using Attention-Based Multiple Instance Learning and Daily Rainfall Data: Case Study Mt. Umyeon, Korea
There are several issues that should be addressed, as follows:
The manuscript is lack in scientific aspect. There is no survey/collection, data process and analysis and just show the existing results. The manuscript looks like report. So, the manuscript is not suitable to published in “Geosciences Journal”.
There is big problem in input data. The salinity, temperature, rainfall and tide data were used for the study. But the station of the data is very few considering the study area. Moreover, the data used after interpolation and this is not accepted because the station is too few. Also, the selection of input data should be justified. To apply the machine learning, the data should be enough. If the data is not enough, the machine learning can’t be used. So, the manuscript is not suitable to published in “Geosciences Journal”.
I can’t find the new findings in data and methodology and the manuscript is not novel. So, the manuscript can’t be published in “Geosciences Journal”
Major Comments
The dataset is too small to support such a broad model comparison. The study uses only 31 annual observations to train and tune 15 models, which creates a serious risk of overfitting, especially for more flexible models such as the 100-unit MLP. The authors should either expand the dataset or narrow the analysis to a smaller set of parsimonious models. In addition, key forecasting baselines are missing. Naïve, drift, linear-trend, exponential-smoothing, ARIMA, and lagged-CO₂ models should be included, because year-only and lagged-CO₂ baselines are essential for determining whether the proposed machine-learning models actually improve upon the historical trend.
The validation and forecasting procedures also require substantial clarification. The current validation strategy is unclear and appears unsuitable for time-series data, since random splitting or shuffled cross-validation may introduce temporal leakage. Model tuning and evaluation should instead rely on rolling-origin or expanding-window validation, preferably within a nested framework. The 2021–2030 forecasts are also not reproducible. Because the models require future values of domestic credit, GDP, population, renewable energy consumption, and trade openness, the manuscript must clearly explain how these future predictor values were generated, forecasted, or assumed under specific scenarios.
Several methodological details and interpretation issues should be addressed before the results can be considered reliable. Preprocessing and hyperparameter tuning are insufficiently described; the manuscript should report scaling procedures, search spaces, validation criteria, random seeds, software versions, and complete model settings. Preprocessing must be fitted only on the training data, and all models should receive comparable tuning effort. The model-selection approach is also inconsistent: although Bagging performs best in Table 5, other models are later selected as the “most plausible” mainly because their forecast curves visually resemble the historical trend. Model selection should follow predefined out-of-sample performance criteria rather than visual judgment. Finally, the XAI analysis is unclear and appears overinterpreted. The manuscript should specify which model was explained using SHAP, LIME, and permutation importance, and given the small sample size and strong correlations among predictors, feature rankings should be treated as unstable predictive associations rather than causal effects.
Integrating Multi-Temporal Synthetic Aperture Radar Observations and Boosting Ensemble Learning for High-Resolution Flood Susceptibility Mapping in Napa County, California
The manuscript has lack in novelty and completeness in many parts. Please check the followings,
Engineering Applications of Artificial Intelligence
An environmental similarity-guided attention neural network for landslide susceptibility mapping in underrepresented environments
After checking the manuscript, the manuscript can’t be published in present form because there are some major issues such as followings,
Geoscience Frontiers
From spatial dependence to temporal memory: A unified graph neural network for spatiotemporal landslide susceptibility prediction
Before the check the model and validation, there are serious missing of basic information such as followings,
So, the manuscript can’t be reviewed and published to the journal.
- The manuscript addresses a relevant topic and the authors have made a clear effort to respond to the previous reviewers. However, I still have a major concern regarding the fundamental design of the study.
- The main issue is that the sediment yield used as the target variable is not based on direct observations. Instead, it was estimated from a discharge–sediment rating relationship. The authors now acknowledge this limitation and the resulting structural dependence between the target and several hydrological predictors.
- This makes the very high machine-learning performance difficult to interpret. In particular, runoff volume, peak discharge, and their derived variables are closely related to the discharge information already used to generate the proxy sediment yield. Therefore, an RF R2 of about 0.94 may partly reflect this built-in relationship rather than true predictive capability.
- The addition of 5-fold cross-validation improves the analysis, but it does not resolve this underlying circularity.
- I am also concerned about the small number of independent events (69) relative to the large number of tested model and feature combinations. Some configurations include up to 196 predictors, which further raises concerns about overfitting and the robustness of the reported results.
- In my view, these limitations affect the central conclusions of the paper and cannot be adequately addressed by further textual revision alone. Independent sediment observations would be necessary to properly validate the proposed framework.