Learning Distributed Parameters of Land Surface Hydrologic Models Using a Generative Adversarial Network

Abstract

Land surface hydrologic models adeptly capture crucial terrestrial processes with a high level of spatial detail. Typically, these models incorporate numerous uncertain, spatially varying parameters, the specification of which can profoundly impact the simulation capabilities. There is a longstanding tradition wherein parameter calibration has served as the conventional procedure to enhance model performance. However, calibrating distributed land surface hydrologic models presents a great challenge, often resulting in uneven spatial performance due to the compression of information inherent in model outputs and observations into a single-value objective function. To address this problem, we propose a novel Generative Adversarial Network-based Parameter Optimization (GAN-PO) method. By leveraging a deep neural network to discern model spatial biases, we train a generative network to produce spatially coherent parameter fields, minimizing distinctions between simulations and observations. By leveraging neural network-based surrogate models to make the physical model differentiable, we employ GAN-PO to calibrate the Variable Infiltration Capacity (VIC) model against evapotranspiration (ET) over China's Huaihe basin. The results show that GAN-PO can diminish errors in simulated ET derived from default parameters across nearly all grid cells within the study region, surpassing the conventional calibration approach based on the parameter regionalization technique. Ablation analysis indicates that relying solely on the traditional loss could lead to deteriorated model performance, underscoring the crucial role of the discriminator. Notably, due to the discriminator's explicit identification of model spatial biases, GAN-PO excels in maintaining spatial consistency, outperforming the state-of-the-art differentiable parameter learning (dPL) method in terms of model spatial performance.