Prediction on start of season (SOS) of ground vegetation phenology incorporating multisource data and neural network

Remote sensing technologies are an effective way to understand vegetation dynamics at regional and global scales in recent decades. However, remote sensing-based vegetation phenology is susceptible to observational errors, spatial scales, and inter-annual climate variations, leading to discrepancies between remote sensing-based and ground vegetation phenology. Meanwhile, traditional phenological and dynamic models, along with climate and land cover types, are significant to influence vegetation phenology prediction. In order to increase prediction accuracy, this study proposes a ground vegetation phenology prediction model that incorporates multi-source data and a deep belief network. The multi-source data include remote sensing images and interdisciplinary factors such as climatology and phenology. To validate the proposed model prediction accuracy, ground vegetation phenology data from North America were selected, and several traditional prediction models were compared. The results show that the proposed model has the highest prediction accuracy, with a CC value of 0.880, a MAE of only 6.065 days, and a RMSE of only 7.990 days. The results demonstrate that the proposed model not only enhances prediction accuracy, but also improves model stability.