A continuous learning framework based on physics-guided deep learning for crop phenology simulation

Abstract

Process-based models (PBMs) and artificial intelligence models (AIMs) are both widely used to simulate crop growth under various environmental conditions and farm management practices. PBMs offer the advantage of interpretable simulations due to their mechanistic underpinnings, but the latest insights from crop growth mechanism research are often not promptly incorporated into PBMs. Further, while AIMs can directly extract potential patterns from data, they struggle to generate temporally continuous simulations due to their lack of consideration for crop growth processes, thus limiting the interpretability of their simulations. To synergize the strengths of PBMs and AIMs, we developed a continuous learning framework, AGLPF (APSIM Guided LSTM Phenology Framework), to dynamically simulate the changes in maize phenology across the Chinese Maize Belt. The AGLPF consists of a PBM (APSIM), its phenology dataset, and an AIM based on attention-Long short-term memory (LSTM). When initially training the AIM in AGLPF by the PBM output dataset, the AGLPF was capable of replicating the PBM outcomes, with an average RMSE of 0.8 days for the vegetative growth phase and flowering phase, 1.4 days for the grain filling phase and 2.0 days for the full growing cycle. With incremental actual phenology data from 0 to 12 years being used for self-tuning training, the simulations of the AGLPF increasingly aligned with actual data. Notably, the RMSE of the full growing cycle steadily declined from 27.8 days to 5.5 days. Moreover, the self-tuning training method performed better than the from-scratch training method in the simulation of all the phenological phases. The development of AGLPF has provided a framework to consider physics-guided AIM to simulate crop phenology and even other crop-related variables while being easy to upgrade and easily interpret outputs.