يمكن أن يحسن التعلم الآلي الموجه بالمعرفة من تقدير دورة الكربون في النظم الزراعية
Knowledge-guided machine learning can improve carbon cycle quantification in agroecosystems

المجلة: Nature Communications، المجلد: 15، العدد: 1
DOI: https://doi.org/10.1038/s41467-023-43860-5
PMID: https://pubmed.ncbi.nlm.nih.gov/38191521
تاريخ النشر: 2024-01-08
المؤلف: Licheng Liu وآخرون
الموضوع الرئيسي: آثار تغير المناخ على الزراعة

نظرة عامة

تناقش هذه القسم أهمية قياس دورة الكربون بدقة في النظم الزراعية للتخفيف من تغير المناخ وتعزيز إنتاج الغذاء المستدام. تواجه أساليب النمذجة التقليدية، سواء كانت قائمة على العمليات أو مدفوعة بالبيانات، عدم يقين كبير في التنبؤ بسبب تعقيد العمليات البيوجيوكيميائية وبيانات المراقبة غير الكافية. لمعالجة هذه التحديات، يقترح المؤلفون إطار عمل تعلم الآلة الموجه بالمعرفة (KGML) الذي يدمج الرؤى من النماذج القائمة على العمليات، وبيانات الاستشعار عن بعد عالية الدقة، وتقنيات تعلم الآلة.

باستخدام حزام الذرة في الولايات المتحدة كدراسة حالة، يظهر إطار KGML أداءً متفوقًا في قياس ديناميات دورة الكربون مقارنة بالطرق التقليدية. ومن الجدير بالذكر أنه يكشف عن تفاصيل مكانية أكثر بنسبة 86% في تغيرات الكربون العضوي في التربة مقارنة بالأساليب التقليدية ذات الدقة المنخفضة. كما يحدد المؤلفون بروتوكولًا لتعزيز KGML، مقترحين إمكانيته في تطوير نماذج هجينة يمكن أن تتنبأ بشكل أفضل بديناميات النظام الأرضي المعقدة. نظرًا لأن النظم الزراعية تشغل حوالي ثلث سطح الأرض وتلعب دورًا حاسمًا في دورة الكربون العالمية، فإن قياس تدفقات الكربون وتغيرات الكربون العضوي في التربة (SOC) بدقة أمر حيوي لاستراتيجيات فعالة للتخفيف من غازات الدفيئة.

الطرق

تحدد قسم “الطرق” في ورقة البحث التصميم التجريبي والتقنيات التحليلية المستخدمة للتحقيق في أسئلة البحث. استخدمت الدراسة نهجًا كميًا، حيث تم دمج التحليلات الإحصائية لتقييم البيانات التي تم جمعها من تجارب مختلفة. شملت المنهجيات المحددة تجارب محكومة، حيث تم التلاعب بالمتغيرات بشكل منهجي لمراقبة تأثيراتها على النتائج المعنية.

شملت جمع البيانات استخدام أدوات وبروتوكولات موحدة لضمان الموثوقية والصلاحية. تم إجراء التحليل باستخدام برامج إحصائية مناسبة، مع تحديد مستويات الدلالة عند p < 0.05. تم تفسير النتائج في سياق الأدبيات الموجودة، مما يسمح بفهم شامل للنتائج وآثارها ضمن المجال. بشكل عام، أسست الدقة المنهجية أساسًا قويًا للاستنتاجات المستخلصة في الدراسة.

النتائج

يمثل إطار KGML-ag-Carbon نهجًا مبتكرًا لمحاكاة الدورات البيوجيوكيميائية في النظم الزراعية من خلال دمج النمذجة القائمة على العمليات مع تقنيات الذكاء الاصطناعي المتقدمة. يستخدم نموذج ecosys المعتمد جيدًا لإبلاغ هيكله، الذي يشمل وحدات فرعية للنباتات، والتربة، وتبادل الكربون الجوي. ميزة رئيسية في KGML-ag-Carbon هي قدرته على استيعاب بيانات الاستشعار عن بعد، وخاصة الإنتاجية الأولية الإجمالية (GPP)، التي تعتبر مدخلًا حيويًا للكربون. يحل النموذج بشكل فعال المكونات الرئيسية لميزانية الكربون، بما في ذلك التنفس الذاتي (Ra)، والتنفس غير الذاتي (Rh)، والتنفس الكلي للنظام البيئي (Reco)، وتبادل الكربون الصافي للنظام البيئي (NEE)، على أساس يومي، بينما يقدر أيضًا العائد السنوي.

لتحسين دقة التنبؤ، خضع KGML-ag-Carbon لمرحلة تدريب مسبق باستخدام أكثر من 14 مليون نقطة بيانات صناعية تم إنشاؤها من نموذج ecosys، مما يقلل بشكل كبير من التكاليف المرتبطة بجمع البيانات في العالم الحقيقي. يتضمن النموذج وظائف خسارة موجهة بالمعرفة (KG) التي تفرض قيودًا بيوجيوكيميائية وفيزيائية، مثل معادلات توازن الكتلة وعوامل الاستجابة، لتحسين توقعاته. تم تحقيق الضبط الدقيق باستخدام مجموعة فرعية من بيانات العائد الحقيقية وقياسات التدفق، مما يدمج البيانات الصناعية والمراقبة للاحتفاظ بمعرفة التدريب المسبق مع ضمان أن تظل مخرجات النموذج استجابة للمدخلات البيئية. لا يحسن هذا النهج الشامل قدرات النموذج التنبؤية فحسب، بل يعالج أيضًا عدم اليقين المتأصل في النمذجة الهجينة.

المناقشة

تقيّم قسم المناقشة في ورقة البحث أداء نموذج KGML-ag-Carbon في التنبؤ بعوائد المحاصيل وتدفقات الكربون، مع تسليط الضوء على فعاليته مقارنة بالنماذج التقليدية. أظهرت التقييمات الأولية أن نموذج KGML-ag-Carbon المدرب مسبقًا حقق اتساقًا عاليًا مع محاكاة ecosys، كما يتضح من قيم R² التي بلغت 0.99 لمقاييس العائد وتدفق الكربون. بعد الضبط الدقيق، أظهر النموذج قيم R² محسنة بلغت 0.91 و0.88 لعوائد الذرة وفول الصويا، على التوالي، و0.94 و0.96 لتنبؤات تدفق الكربون اليومي. تم التحقق من متانة النموذج من خلال اختبارات مختلفة، مما يشير إلى أداء متفوق على نماذج تعلم الآلة البحتة، خاصة في السيناريوهات التي تحتوي على بيانات تدريب محدودة. تأتي مزايا KGML-ag-Carbon من تدريبه المسبق باستخدام بيانات صناعية وهيكل مخصص يدمج المعرفة العلمية، مما يسمح بتنبؤات موثوقة حتى مع عدد أقل من العينات المعلّمة.

بالإضافة إلى ذلك، تناقش القسم طرق تقليل عدم اليقين في توقعات KGML-ag-Carbon، مع التأكيد على أهمية دمج بيانات الإنتاجية الأولية الإجمالية (GPP) والتدريب المسبق باستخدام مجموعات بيانات صناعية. تعزز هذه الاستراتيجيات بشكل كبير أداء النموذج، خاصة في الحالات التي تفتقر إلى البيانات. كما تساهم الهيكلية الهرمية للنموذج ووظائف الخسارة الموجهة بالمعرفة في التقاط ديناميات الكربون المعقدة، على الرغم من أن تأثيرها أقل وضوحًا مقارنة بمدخلات GPP والتدريب المسبق. تظهر التنبؤات عالية الدقة لتدفقات الكربون وعوائد المحاصيل عبر الغرب الأوسط الأمريكي إمكانيات النموذج في قياس ميزانية الكربون الإقليمية بدقة، مما يوفر نهجًا فعالًا من حيث التكلفة لتحسين تقييمات احتجاز الكربون. بشكل عام، يمثل KGML-ag-Carbon تقدمًا كبيرًا في نمذجة الكربون الزراعي، مع آثار على تحسين ممارسات الإدارة وزيادة دقة تقديرات غازات الدفيئة.

Journal: Nature Communications, Volume: 15, Issue: 1
DOI: https://doi.org/10.1038/s41467-023-43860-5
PMID: https://pubmed.ncbi.nlm.nih.gov/38191521
Publication Date: 2024-01-08
Author(s): Licheng Liu et al.
Primary Topic: Climate change impacts on agriculture

Overview

The section discusses the importance of accurately quantifying the carbon cycle in agroecosystems to mitigate climate change and promote sustainable food production. Traditional modeling approaches, whether process-based or data-driven, face significant prediction uncertainties due to the complexity of biogeochemical processes and insufficient observational data. To address these challenges, the authors propose a Knowledge-Guided Machine Learning (KGML) framework that integrates insights from process-based models, high-resolution remote sensing data, and machine learning techniques.

Using the U.S. Corn Belt as a case study, the KGML framework demonstrates superior performance in quantifying carbon cycle dynamics compared to conventional methods. Notably, it reveals 86% more spatial detail in soil organic carbon changes than traditional coarse-resolution approaches. The authors also outline a protocol for enhancing KGML, suggesting its potential for developing hybrid models that can better predict complex earth system dynamics. Given that agroecosystems occupy approximately one-third of the Earth’s land surface and play a crucial role in the global carbon cycle, accurately quantifying carbon fluxes and changes in soil organic carbon (SOC) is vital for effective greenhouse gas mitigation strategies.

Methods

The “Methods” section of the research paper outlines the experimental design and analytical techniques employed to investigate the research questions. The study utilized a quantitative approach, incorporating statistical analyses to evaluate the data collected from various experiments. Specific methodologies included controlled experiments, where variables were systematically manipulated to observe their effects on the outcomes of interest.

Data collection involved the use of standardized instruments and protocols to ensure reliability and validity. The analysis was conducted using appropriate statistical software, with significance levels set at p < 0.05. The results were interpreted in the context of existing literature, allowing for a comprehensive understanding of the findings and their implications within the field. Overall, the methodological rigor established a solid foundation for the conclusions drawn in the study.

Results

The KGML-ag-Carbon framework represents an innovative approach to simulating biogeochemical cycles in agroecosystems by integrating process-based modeling with advanced artificial intelligence techniques. It utilizes the well-validated ecosys model to inform its architecture, which encompasses submodules for plants, soil, and atmospheric carbon exchange. A key feature of KGML-ag-Carbon is its ability to assimilate remote sensing data, particularly Gross Primary Production (GPP), which serves as a critical carbon input. The model effectively resolves major components of the carbon budget, including autotrophic respiration (Ra), heterotrophic respiration (Rh), total ecosystem respiration (Reco), and net ecosystem carbon exchange (NEE), on a daily basis, while also estimating annual yield.

To enhance prediction accuracy, KGML-ag-Carbon underwent a pre-training phase using over 14 million synthetic data points generated from the ecosys model, which significantly reduces the costs associated with real-world data collection. The model incorporates knowledge-guided (KG) loss functions that impose biogeochemical and physical constraints, such as mass balance equations and response thresholds, to refine its predictions. Fine-tuning was achieved using a subset of real-world yield data and flux measurements, merging synthetic and observational data to retain pre-training knowledge while ensuring that the model’s outputs remain responsive to environmental inputs. This comprehensive approach not only improves the model’s predictive capabilities but also addresses uncertainties inherent in hybrid modeling.

Discussion

The discussion section of the research paper evaluates the performance of the KGML-ag-Carbon model in predicting crop yields and carbon fluxes, highlighting its effectiveness compared to traditional models. Initial assessments showed that the pre-trained KGML-ag-Carbon model achieved high consistency with ecosys simulations, evidenced by R² values of 0.99 for yield and carbon flux metrics. After fine-tuning, the model demonstrated improved R² values of 0.91 and 0.88 for corn and soybean yields, respectively, and 0.94 and 0.96 for daily carbon flux predictions. The model’s robustness was further validated through various tests, indicating superior performance over pure machine learning models, particularly in scenarios with limited training data. The advantages of KGML-ag-Carbon stem from its pre-training with synthetic data and a customized structure that integrates scientific knowledge, allowing for reliable predictions even with fewer labeled samples.

Additionally, the section discusses pathways to reduce uncertainty in KGML-ag-Carbon’s predictions, emphasizing the importance of incorporating Gross Primary Production (GPP) data and pre-training with synthetic datasets. These strategies significantly enhance model performance, especially in data-scarce situations. The model’s hierarchical structure and knowledge-guided loss functions also contribute to capturing complex carbon dynamics, although their impact is less pronounced compared to GPP inputs and pre-training. The high-resolution predictions of carbon fluxes and crop yields across the U.S. Midwest demonstrate the model’s potential for precise regional carbon budget quantification, offering a cost-effective approach to improving carbon sequestration assessments. Overall, KGML-ag-Carbon represents a significant advancement in agricultural carbon modeling, with implications for optimizing management practices and enhancing the accuracy of greenhouse gas estimations.