Emulation of Dynamic Process-Based Agroecosystem Models Using Long Short-Term Memory Networks
Aakula, Viivi (2024-04-09)
Emulation of Dynamic Process-Based Agroecosystem Models Using Long Short-Term Memory Networks
(09.04.2024)
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
avoin
Julkaisun pysyvä osoite on:
https://urn.fi/URN:NBN:fi-fe2024042220475
https://urn.fi/URN:NBN:fi-fe2024042220475
Tiivistelmä
Modeling carbon balance in agroecosystems help monitoring changes in carbon emissions and influence in ecosystem functioning and productivity. Process-based models are widely used in modeling diverse agroecosystems, and also enable quantification of carbon balance in agroecosystems. However, process-based models tend to be very computationally demanding, due to their complex computations based on hypotheses and assumption of the dynamics of the system. The computational demands complicate performing large scale simulations, needed when simulating several different parameter scenarios, such as model calibration and sensitivity analysis.
In order to mitigate the computational burden of large scale simulations, a surrogate model utilizing neural networks is developed to emulate the behavior of a process-based land model BASGRA\_N, obtaining a fast execution time. The emulator recognizes sequentially dependent data by networks specifically designed for sequential learning. Additionally, it is applicable to other similar agroecosystem models. The model is evaluated by 5-fold cross validation, achieving RMSEs of 0.0290 (g C m^(-2) h^(-1)) and 0.322 (m^2 m^(-2)) for weekly mean values of hourly NPP and LAI, respectively. Each of the 5 folds give R^2 of >0.91 for NPP and >0.93 for LAI.
The thesis begins with basic concepts on neural networks, concerning to regression tasks, covering a fundamental neural network model, its architecture, features, and general training methods. Subsequently, the study continues to sequential modeling and introduces neural networks designed for processing sequentially structured data. Subsequently, an overall review on existing research on machine learning applications, especially in emulation of process-based models, is provided. Lastly a novel emulator model applying neural networks is introduced for emulation of an agroecosystem model.
This project was done in collaboration with Carbon Cycle group of Finnish Meteorological Institute, for their requirement for an emulator for a process-based agroecosystem model BASGRA_N to enable large scale simulations for simulator calibration purposes.
In order to mitigate the computational burden of large scale simulations, a surrogate model utilizing neural networks is developed to emulate the behavior of a process-based land model BASGRA\_N, obtaining a fast execution time. The emulator recognizes sequentially dependent data by networks specifically designed for sequential learning. Additionally, it is applicable to other similar agroecosystem models. The model is evaluated by 5-fold cross validation, achieving RMSEs of 0.0290 (g C m^(-2) h^(-1)) and 0.322 (m^2 m^(-2)) for weekly mean values of hourly NPP and LAI, respectively. Each of the 5 folds give R^2 of >0.91 for NPP and >0.93 for LAI.
The thesis begins with basic concepts on neural networks, concerning to regression tasks, covering a fundamental neural network model, its architecture, features, and general training methods. Subsequently, the study continues to sequential modeling and introduces neural networks designed for processing sequentially structured data. Subsequently, an overall review on existing research on machine learning applications, especially in emulation of process-based models, is provided. Lastly a novel emulator model applying neural networks is introduced for emulation of an agroecosystem model.
This project was done in collaboration with Carbon Cycle group of Finnish Meteorological Institute, for their requirement for an emulator for a process-based agroecosystem model BASGRA_N to enable large scale simulations for simulator calibration purposes.