2:30 PM - 2:45 PM
[AHW22-04] Calibrating Land Surface Model Parameters for Crop Canopy Energy Cycle Using Differentiable Parameter Learning
Keywords:Differentiable Parameter Learning, Land surface model, energy cycle
Environmental changes, particularly climate warming and accelerated urbanization, underscore the importance of understanding land dynamics, especially for agricultural production. Land Surface Models (LSMs) simulate the interactions between the Earth's land surface and atmosphere, providing detailed insights into multiple ecosystem dynamics. A central challenge with LSMs is the strong influence of unobservable or underdetermined parameters on their behavior and skill. In the agricultural domain, the accuracy of crop growth and yield simulations relies heavily on the setting of vegetation type parameters, such as the maximum water storage capacity of the vegetation layer. Parameter calibration has been a fundamental practice in various geoscientific domains for decades. Despite the introduction of numerous methods such as Parameter Ensembles, Monte Carlo algorithms, and Evolutionary Algorithms, these approaches are often characterized by high computational costs, susceptibility to local optima, and slow convergence rates. Additionally, their reliance on random optimization strategies contributes to their inefficiency. Although some parameter transformation methods exist, their structures are based on human cognition, which can rigidly constrain the effectiveness of parameter information. This research proposes a novel approach leveraging deep learning techniques to identify optimal parameters and enhance parameterization efficiency for agricultural productivity. By creating a unified parameter set, we aim to improve the accuracy of LSMs, leading to better decision-making and resource management.