Aboveground biomass change of the alpine meadow on the Qinghai-Tibet Plateau under climate warming

Abstract

Livestock grazing and pika (Och-otona curzoniae) disturbance are two key determinants of the functionality of the grazing-adapted alpine grassland on the Qinghai-Tibet Plateau (QTP). Overgrazing and rodent outbreaks have decreased grassland productivity and degraded this fragile ecosystem. Grassland aboveground biomass (AGB) is a reliable indicator of grassland health and ecosystem service value. It provides a quantitative measure of forage production and can be used to assess livestock carrying capacity (LCC). Monitoring AGB in a timely and accurate manner across multiple scales to track spatiotemporal variability in AGB is a prerequisite of formulating better sustainable development strategies, especially in climate sensitive areas such as the QTP. This doctoral thesis develops advanced AGB modeling methods to evaluate the effects of the two aforementioned external disturbances on AGB at two scales (micro and watershed) and assess the spatiotemporal variability of AGB and evaluate the sustainable LCC at the watershed scale in light of climate warming. Comprising four main chapters, Chapters 3 and 5 develop the AGB modeling methodology and lay the foundation of the analysis of spatiotemporal AGB change for Chapter 4 at the micro (plot)-scale and Chapter 6 at the watershed scale, respectively. Chapter 3 addresses how to improve AGB estimation from unpiloted aerial vehicle (UAV) images by taking into account the effect of varied external disturbance severities (mowing-simulated grazing and number of pika). Ground truth data were supplemented by information on bare patches extracted from UAV imagery. The estimation accuracy of the model (R2 ) increased from 0.44 to 0.81 compared with the initial model based solely on a vegetation index. However, this modification overestimated bare ground AGB by about 19-38 g m−2. Drawing from the modeling results in Chapter 3, Chapter 4 analyzed the precise effect of external disturbances on AGB. In total, three levels of grazing (simulated via mowing) and three levels of pika disturbance were carried out over a period of three years. This chapter reported that the effects of pika disturbance on AGB change were overwhelmed by the significantly different AGB at different mowing severities (-0.471 < r < -0.368), but can still be identified by inspecting each mowing intensity (-0.884 < r < -0.626). The impact of severe mowing on AGB loss was more profound than that of severe pika disturbance in heavily disturbed plots, and the joint effects of both severe disturbances had the most impact on AGB loss. However, pika disturbance made little difference to AGB change in the moderate and non-mowed plots. Mowing intensity weakens the relationship between pika population and AGB change, but pika disturbance hardly affects the relationship between mowing severity and AGB change. Results indicate a reducing grazing intensity simulated via mowing was more effective than controlling pika population to achieve sustainable grazing of heavily disturbed grassland. Chapter 5 researched how to quantify AGB at the watershed scale from various fine- to medium-resolution satellite images using machine learning methods, incorporating the findings from Chapter 4 to improve the estimation accuracy. It also develops a new framework for AGB modeling and mapping from a variety of environmental factors. The effectiveness of four widely used models with various input combinations was assessed in estimating AGB based on Google Earth Engine and Tensorflow. This chapter also assessed the importance of the considered factors in the modeling. The proposed modeling framework successfully assisted machine leaning regression models to map the spatial distribution of AGB. Categorical variables, including vegetation types (grass and shrub) and observation time, improved the accuracy (R2) by 0.112-0.216 for machine learning models. Grass AGB was less accurately predicted than shrub AGB, but the pooling of both vegetation types improved estimation accuracy by 0.171-0.269. The deep neural network model had the highest accuracy of 0.818 using all non-field measured variables as the inputs. Chapter 6 builds on the estimation models developed in Chapter 5 to map the spatiotemporal variability in AGB and analyzed its values under three climate warming scenarios to derive the theoretical LCC, taking into account the effects of pika disturbance determined in Chapter 4. Results show that elevation and slope are the key determinants of the distribution of AGB and its change. Although grasses and shrubs occupied an equivalent area, the latter generated approximately double the amount of total AGB, but the former generated slightly more available AGB. Although pika habitat extended over 45.3% of the study area, pika only reduced the total available AGB by less than 0.07%. The actual LCC of 2.16 SU/ha (sheep unit per hectare) of the study area was lower than the maximum theoretical LCC of 2.57-2.80 SU/ha, which can be increased by 0.95-1.40, 1.30-1.65 and 0.9-1.35 SU/ha in light of climate warming scenarios of 1.0, 1.5 and 2.0℃, respectively. Although systematic climate warming increased AGB, its impacts were especially marked at lower elevations and on north-facing slopes because of higher soil water availability. Used effectively, these models can be used to track local scale and watershed scale variability to understand the susceptibility of AGB dynamics elsewhere, thereby underpinning more systematic and precise determinations of LCC and grazing practices. This study can be improved by collecting more field samples other than 1,469 used over a longer period. It will be interesting to see whether similar findings can be obtained from another catchment. The comparison of results from multiple catchments can help to assess how environmental heterogeneity affects the calculated theoretical LCC.