التنبؤ العالمي بالفيضانات الشديدة في الأحواض المائية غير المقاسة
Global prediction of extreme floods in ungauged watersheds

المجلة: Nature، المجلد: 627، العدد: 8004
DOI: https://doi.org/10.1038/s41586-024-07145-1
PMID: https://pubmed.ncbi.nlm.nih.gov/38509278
تاريخ النشر: 2024-03-20
المؤلف: Grey Nearing وآخرون
الموضوع الرئيسي: دراسات الهيدرولوجيا وإدارة أحواض المياه

نقاش

تسلط قسم النقاش في ورقة البحث الضوء على التقدم الذي حققه نموذج الذكاء الاصطناعي باستخدام شبكات الذاكرة طويلة وقصيرة المدى (LSTM) في التنبؤ بتدفق المياه اليومي على مدى فترة توقع تبلغ 7 أيام. تم تدريب النموذج والتحقق من صحته بدقة باستخدام بيانات خارج العينة من 5,680 مقياس تدفق، مع استخدام التحقق المتقاطع العشوائي k-fold لضمان مقاييس أداء قوية. ومن الجدير بالذكر أن نموذج الذكاء الاصطناعي أظهر موثوقية محسنة في التنبؤ بالأحداث الهيدرولوجية القصوى مقارنة بمعيار GloFAS، محققًا درجات دقة واسترجاع أعلى عبر فترات عودة مختلفة. على سبيل المثال، تفوق نموذج الذكاء الاصطناعي على GloFAS في 64% إلى 76% من المقاييس لفترات العودة التي تتراوح من 1 إلى 10 سنوات، مع تحسينات ذات دلالة إحصائية واضحة بشكل خاص للأحداث التي تستمر لمدة عامين وخمسة أعوام.

تظهر التحليلات أيضًا أنه بينما يظهر نموذج الذكاء الاصطناعي عمومًا أداءً متفوقًا، فإن موثوقية التنبؤات تختلف جغرافيًا، مع اختلافات ملحوظة في درجات F1 عبر القارات. كانت فعالية نموذج الذكاء الاصطناعي ملحوظة بشكل خاص في جنوب غرب المحيط الهادئ، بينما أظهر GloFAS موثوقية أقل في أمريكا الجنوبية. بالإضافة إلى ذلك، تتناول الدراسة التحدي المتمثل في التنبؤ بأداء النموذج في الأحواض غير المقاسة، مشيرة إلى أنه بينما يكون نموذج الذكاء الاصطناعي غالبًا أكثر موثوقية، فإن الأنماط النظامية لتفضيل النموذج بناءً على خصائص الحوض لا تزال غير واضحة. تؤكد النتائج على إمكانيات الأساليب المدفوعة بالذكاء الاصطناعي لتعزيز التنبؤ بالفيضانات وأنظمة الإنذار المبكر، مما يبرز الحاجة إلى زيادة الوصول إلى البيانات الهيدرولوجية لتحسين التنبؤات العالمية للفيضانات.

Journal: Nature, Volume: 627, Issue: 8004
DOI: https://doi.org/10.1038/s41586-024-07145-1
PMID: https://pubmed.ncbi.nlm.nih.gov/38509278
Publication Date: 2024-03-20
Author(s): Grey Nearing et al.
Primary Topic: Hydrology and Watershed Management Studies

Discussion

The discussion section of the research paper highlights the advancements made by an AI model utilizing long short-term memory (LSTM) networks for predicting daily streamflow over a 7-day forecast horizon. The model was rigorously trained and validated using out-of-sample data from 5,680 streamflow gauges, employing random k-fold cross-validation to ensure robust performance metrics. Notably, the AI model demonstrated improved reliability in forecasting extreme hydrological events compared to the GloFAS benchmark, achieving higher precision and recall scores across various return periods. For instance, the AI model outperformed GloFAS in 64% to 76% of gauges for return periods ranging from 1 to 10 years, with statistically significant improvements particularly evident for 2-year and 5-year events.

The analysis further reveals that while the AI model generally exhibits superior performance, the reliability of forecasts varies geographically, with notable differences in F1 scores across continents. The AI model’s effectiveness was particularly pronounced in the Southwest Pacific, whereas GloFAS showed lower reliability in South America. Additionally, the study addresses the challenge of predicting model performance in ungauged basins, indicating that while the AI model is often more reliable, systematic patterns for model preference based on catchment attributes remain elusive. The findings underscore the potential of AI-driven approaches to enhance flood forecasting and early warning systems, emphasizing the need for increased access to hydrological data to further improve global flood predictions.