Embedded physical constraints in machine learning to enhance vegetation phenology prediction

Abstract

Vegetation phenology plays a pivotal role in ecological processes on terrestrial surfaces and the interactions between the biosphere and atmospheric feedback. Current attempts to retrieve vegetation phenology have primarily depended on vegetation indices extracted from satellite remote sensing imagery. These approaches often neglect the underlying physical mechanisms associated with climatic factors, and there is a notable absence of evaluations and comparisons with field-observed phenology inventory data. To address these limitations, this paper proposes an innovative physical constraint neural networks (PCNNs) model that combines machine learning techniques with physical mechanisms to enhance the accuracy of vegetation phenology predictions. By incorporating meteorological variables into a machine learning model and by using the Moderate-Resolution Imaging Spectroradiometer (MODIS) dataset to identify the vegetation phenology of four vegetation types in North America, this study delved into the relationship between vegetation phenology and climate factors as well as its impacts on ecosystems. Our model demonstrated high accuracy compared to machine learning methods without physical mechanisms when validated by field observations from PhenoCam and the USA National Phenology Network (USA-NPN) spanning from 2001 to 2021. The results show that the overall root mean square error (RMSE) with physical constraints is reduced to 12.37 days, higher by 2.6 days than the machine learning method without physical constraints. We compared four vegetation types using different machine learning and traditional rule-based methods, deciduous vegetation (DV) exhibited the most favorable prediction results, with the RMSE and mean bias error (MBE) as low as 5.71 days and 4.06 days in PCNNs model, respectively. This was followed by evergreen needle-leaved forests and mixed forests with RMSE of 12.32 and 13.28 days, respectively. The stressed deciduous vegetation type had the worst prediction result of 19.86 days (RMSE), and the weighted index of agreement (WIA) attained a value of 0.68. These findings suggest that the model embedded with physical constraints significantly boosted prediction accuracy for four common vegetation types, particularly for DV, compared to the unconstrained ML model. It offers valuable insights into the incorporation of physical mechanisms within machine learning models. This research paves the way for substantial advancements in the field of land surface phenology, enabling more accurate and reliable predictions of vegetation phenology in various ecological and climatic contexts.

Keywords

• Vegetation phenology
• physical mechanisms
• machine learning
• land surface phenology
• MODIS
• North America