تقليل الإفراط في التكيف وعدم الاتساق الفيزيائي في رسم خرائط العرض للفيضانات: إطار تعلم آلي تطوري مقيد بالفيزياء للحمّانات الألبية غير المقاسة

Key Points

• The research aims to improve flood susceptibility mapping by combining physical models with machine learning frameworks.
• Developed a Physically constrained Particle Swarm Optimization–Random Forest (P-PDRF) framework.
• Validated the framework using the Lhasa River Basin as a case study.
• Utilized a maximum daily runoff depth as a ‘Relative Hydraulic Intensity Index’ for model integration.
• Introduced a Physiographically Constrained Negative Sampling (PCNS) strategy.
• Implemented a PSO-optimized ‘Shallow Tree’ configuration to enhance model performance.
• The P-PDRF framework achieved an AUC of 0.942, outperforming standard machine learning models.
• Empirical results indicated that the dynamic index is more crucial than the static Topographic Wetness Index.
• The new approach effectively corrected errors in identifying high-risk flood zones.

Abstract