Spatially structured flood samples and environmental variables may overestimate predictive performance when nearby samples are shared

The gap

Evidence profile

Sourced from the future work and stated challenges of the source papers, classified as general, spanning 3 journals.

Supporting evidence — 3 representative gaps

• A new machine learning technique for predicting river water quality using AVOA-RNN (2026) · Water Science · doi

  This study proposed a hybrid AVOA-RNN framework for predicting river water quality using the Cauvery River data- set. By integrating advanced imbalance-handling strategies (SMOTE, SMOGN) with African Vulture Optimization for hyper-parameter tuning, the model demonstrated supe- rior predictive capability in both regression and classifica- tion tasks. In classification, the proposed model achieved 97.3% accuracy, 99.52% sensitivity, 97.42% specificity, and a ROC-AUC of 0.97, outperforming comparative approaches such as CNN (93.85%), LSTM (94.92%), RNN (95.73%), GRU (96.38%), WOA-NN (0.82), GA-LSTM (0.80), EMD- (F1 = 0.86), WOA-LSTM (F1 = 0.84), AQPSO-SOFNN Y. et al. Water Science (2026) 40:16 Page 19 of 21 Fig. 11 SHAP-based feature importance (dissolved oxygen (DO ≈ 1.3), BOD (≈ 0.65), pH (≈ 0.45), and TDS (≈ 0.25) dominate, while NO₃, Cl, PO₄, SO₄, turbidity, and FC contribute marginally) Fig. 12 Comparative SHAP analysis for the existing models Y. et al. Water Science (2026) 40:16 Page 20 of 21 and SOFNN-HPS (Accuracy = 0.87). In regression tasks, AVOA-RNN obtained the lowest RMSE (8.3) compared to WOA-NN (9.0), GA-LSTM (9.6), EMD-WOA-LSTM (8.8), AQPSO-SOFNN (8.7), and SOFNN-HPS (8.6), along with a high coefficient of determination ( R2 =0.91). These results confirm that AVOA-RNN not only surpasses tradi- tional and hybrid baselines but also ensures greater ecolog- ical interpretability through SHAP-based feature analysis, highlighting DO, BOD, pH, and TDS as dominant param- eters. The framework thus provides a robust and interpret- able solution for sustainable water quality monitoring and management.

  generalfuture work
  Keywords: lstm water sofnn avoa shap proposed hybrid framework river quality model regression tasks accuracy comparative
• Evaluation of flood susceptibility through an artificial neural network-based differential evolution optimization algorithms and GIS techniques (2026) · Natural Hazards · doi

  In this study, the flood susceptibility of the Putna River basin was assessed using advanced machine learning models, namely MLP and three optimized hybrid approaches: ABC–MLP, EHO–MLP and DE–MLP, integrated in a GIS environment. The combination of these mod- Natural Hazards (2026) 122:4061 3 406 Page 28 of 32 els for the first time in the literature for flood susceptibility estimation and their integration in the GIS environment represents the main novelty of the present research work. The comparative analysis clearly demonstrated the superiority of the hybrid models over the classic MLP neural network, highlighting the essential role of optimization algorithms in increasing predictive performance. Among the tested models, DE–MLP achieved the best results, recording an accuracy of 0.982 and an AUC-ROC value of 0.985 on the test set, indicating an excellent discrimi- nation capacity between flood-prone and unaffected areas. Also, the high values of preci- sion, recall and F1 score (all 0.982) confirm the balanced and robust nature of this model. The EHO–MLP and ABC–MLP models performed very well, with accuracies of 0.964 and 0.946, respectively, and AUC-ROC values of 0.970 and 0.975, but lower than the DE–MLP model. The simple MLP model performed the worst (AUC-ROC = 0.945), highlighting the clear advantage of hybrid approaches. The implications of this work also explain the better flood management with the opti- mized models produce flood susceptibility maps that can be used for better flood manage- ment. High risk areas can be targeted for flood mitigation, reducing impact on communities and infrastructure. And finally, accurate flood susceptibility assessment means better resource allocation, focusing on high risk areas. The analysis of the importance of variables indicated slope (0.334), altitude (0.249), distance from the river (0.244) and rainfall (0.183) as the dominant factors in controlling the flooding processes. The final susceptibility maps highlight the high-risk areas, providing valuable support for territorial planning, flood risk management and decision-making at local and regional levels. The results confirm that the DE–MLP model represents an efficient and reliable solution for assessing flood susceptibil- ity in regions with similar characteristics. Based on the new findings of this work, as well as its limitations, the future works can consider other areas to explore with different geography and climate to test the models. In addition, real time data such as rainfall and river flow for dynamic flood susceptibility assessment is an interesting problem and they may assist improve the flood susceptibil- ity mapping works. Combine multiple machine learning algorithms to create more accu- rate models is also a great work that can build on this research and improve flood risk management. Acknowledgements This work was supported by the Henan Science and Technology Research Project (Grant No. 262102321015), the Key Scientific Research Project of Colleges and Universities of Henan Provincial Department of Education (Grant No. 26A410007) and the National Scientific Research Project Cultivation Fund of Huanghuai University (Grant No. XKPY-2023023).The author Romulus Costache acknowledges the financial support received from the Research Institute of the University of Bucharest (ICUB) under the research fellowship programme Fellowship for Young Researchers, Contract no. 12949/12.12.2025. Author contribution Ziguang He – Revisions, Methodology, Software, Conceptualization, Qiancheng Fang: Writing – original draft, Visualization, Methodology, Software, Conceptualization, Funding acquisition, Writing – original draft. Romulus Costache: Writing – review & editing, Conceptualization, Writing – origi- nal draft. Data availability The authors do not have permission to share data. Natural Hazards (2026) 122:4061 3Declarations Competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Page 29 of 32 406

  generalfuture work
  Keywords: flood models susceptibility areas risk high model writing river hybrid better management project grant conceptualization
• Public-service exposure-oriented coastal flood susceptibility assessment and priority zone identification in the Pearl River Estuary (2026) · Frontiers in Marine Science · doi

  Spatially structured flood samples and environmental variables may overestimate predictive performance when nearby samples are shared between training and testing subsets. The need to distinguish flood-prone environmental conditions from exposed receptors. The challenge of interpreting machine-learning models to identify key drivers of coastal flood susceptibility.

  generalstated challengesevidence 5/5
  Keywords: spatially structured flood samples environmental variables overestimate predictive

Questions about this gap