Multi-factor Impact Analysis of Grassland Phenology Changes on the Qinghai-Xizang Plateau Based on Interpretable Machine Learning

Abstract

The vegetation phenology of the Qinghai-Xizang Plateau is changing significantly in the context of climate change. However, there are many hydrothermal factors affecting the phenology, and few studies have focused on the effects of multiple factors on the phenology of the Qinghai-Xizang Plateau, resulting in a lack of understanding of the mechanisms underlying phenological changes on the Qinghai-Xizang Plateau. In this study, we used remote sensing data interpretation to analyze the spatial and temporal variability of grassland phenology on the Qinghai-Xizang Plateau from 2002 to 2021, focusing on precipitation, temperature, altitude, soil, and other aspects to reveal the dominant factors of phenological variability using an interpretable machine learning method (SHAP) and to quantify the interactive effects of multiple factors on phenology. The results showed that:① The growing season start (SOS) of grasslands on the Qinghai-Xizang Plateau mostly ranged from 110 to 150 d, with 56.32 % of grasslands showing an early SOS trend; the growing season end (EOS) mostly ranged from 290-320 d, with 67.65 % of grasslands showing a delayed EOS trend; and the growing season length (LOS) mostly ranged from 120 to 210 d, with 65.50 % of the grasslands showing a trend towards longer growing season lengths. ② SOS in grasslands on the Qinghai-Xizang Plateau was mainly influenced by moisture conditions, in which soil moisture between 10 and 25 kg·m-2 in the 0-10 cm soil layer in March promoted the advancement of SOS and peaked at approximately 20 kg·m-2. EOS was mainly influenced by temperature, with higher temperatures in September and October having a stronger effect on EOS latency promotion and peaking at over 8 ℃ and -0.5 ℃, respectively. The main influencing factors of LOS were more consistent with SOS, in which soil moisture between 15 and 25 kg·m-2 in the 0-10 cm soil layer in March promoted the prolongation of LOS and peaked at approximately 18 kg·m-2. ③ There was an obvious interactive effect of water and heat and other factors on phenology; after soil moisture reached 20 kg·m-2 in the 0-10 cm soil layer in March, SOS was more advanced in low-precipitation and low-altitude areas. Better moisture conditions were more conducive to EOS delay at temperatures above 0 ℃ in October, and soil moisture in high precipitation areas promoted LOS prolongation more when soil moisture was between 12 and 22 kg·m-2 in 0-10 cm in March. The results also demonstrated that interpretable machine learning methods could provide a new approach to the analysis of the multifactorial effects of phenological change.