Parameter identification based on statistical and neural network approaches for the vegetation-water model.

Turing patterns emerging from the vegetation-water model exhibit complex spatial and networked structures, while parameter identification of these patterns has become a challenging inverse problem. This paper aims to present two types of methods for parameter identification, based on a vegetation-water model coupled with climate data on precipitation, temperature, and carbon dioxide concentration in Zhangye. The statistical approach identifies parameters through handcrafted image feature matching using the distance metric. In addition, the deep learning method is employed for parameter identification, one is the modified ResNet50 with a regression head and integrated regularization to enhance generalization; the other is the improved VGG19 that adopts the Gaussian Error Linear Unit (GELU) function and mixed-precision training for greater efficiency. The identification results show that the deep learning methods achieve superior accuracy and robustness compared to the statistical approach, and ResNet50 achieves the best overall performance. Normalized difference vegetation index (NDVI) data further validate the numerical simulation results. Results from parameter identification on patterns enhance the parameterization and predictive capacity of vegetation-water models under climate change.