Improving the reliability of vegetation phenology detection from satellite time-series data

Abstract

Vegetation phenology, i.e., the seasonal cycle of vegetation growth, is one of the most sensitive indicators of climate change effects on terrestrial ecosystems. Satellite remotely sensed images provide the capability for large-scale and long-term vegetation phenology monitoring. In recent decades, an increasing number of studies used satellite-derived vegetation index (VI) time series to extract vegetation phenological metrics, but it may cause phenological results to be misestimated if the satellite images and technologies are inappropriately used. In this thesis, I aim to improve the reliability of vegetation phenology detection from satellite time-series data by addressing several questions: (1) How does satellite spatial resolution affect vegetation spring phenology detection? (2) How does the smoothing process of satellite-derived VI time series affect vegetation spring phenology detection? (3) Are the vegetation phenology monitoring through fusing multi-source satellite images feasible? (4) What is the mechanism of urban-induced microclimate effects on winter wheat growth based on reliable satellite time-series data? To address question 1, I investigated the impact of spatial resolution on the rural-urban difference of vegetation spring phenology using satellite images at different spatial resolutions. The results reveal that vegetation spring phenology in urban areas happen earlier than in rural areas no matter which spatial resolution from 10 m to 8 km is used, but the rural-urban difference in vegetation spring phenology is amplified with spatial resolution, i.e., coarse-resolution satellite images overestimate the urbanization effects on vegetation spring phenology. I further explored that the underlying reason for this overestimation is that the majority of urban pixels in coarser images have higher diversity in terms of spring phenology dates, which leads to spring phenology detected from coarser satellite images earlier than the actual dates. To address question 2, I optimized the parameters of the smoothing process of satellite-derived VI time series based on local real cloud to produce the most accurate spring phenological metrics. The optimal parameters for both the maximum value composite (MVC) and smoothing filters, i.e., iterative Savitzky-Golay (SG) and penalized cubic splines (SP) filters, showed significant spatial heterogeneity. Furthermore, validation with ground PhenoCam data indicated that optimal parameters of the MVC and smoothing filters can produce more accurate results than official vegetation phenology products that use uniform parameters. Specifically, the R2 values of the NASA product (VNP22Q2) and the USGS product (MCD12Q2) were 0.58 and 0.67, which were increased to 0.70 and 0.81, respectively, by the optimal smoothing process. Last, I provide the optimal parameters of the MVC and smoothing filters in each 5° × 5° sub-region over the northern hemisphere (north of 30 °N), which may help future studies to improve the accuracy of phenology detection from satellite-derived VI time series. To address question 3, I assessed the feasibility of two spatiotemporal data fusion algorithms (pair-based: STARFM and time-series-based: SSFIT) on capturing the spatial and temporal variations of spring phenology at fine scales. Compared with results extracted by coarse images, data fusion algorithms generally can produce more accurate spring phenological results at fine scales in both spatial and temporal dimensions. However, the sensitivity of two data fusion algorithms to input data is inconsistent, i.e., the accuracy of STARFM spring phenology depends on the number and temporal distribution of input data while SSFIT is insensitive to them. In addition, STARFM has better performance on capturing spatial and temporal variations of spring phenology than that of SSFIT except when the number of annual input data is less than 6 images and severe cloud-coverage scenario. Even so, the contributions of data fusion algorithms to fine-resolution images decrease when the number of input data is sufficient (e.g., mild cloud coverage). To address question 4, I investigated the urban-induced microclimate effects on winter wheat spring phenology (i.e., the regreen-up date, RGUD) in three cities spanning a range of sizes in northern China, Shijiazhuang, Baoding, and Linqing. The RGUD shows a significant increasing trend along the urban-rural gradients in both Shijiazhuang and Baoding, suggesting that urban-induced increases in temperature indeed advance the spring phenology of winter wheat. Moreover, the maximum influence size of the urban-induced temperature effects on the RGUD is positively correlated with city size, i.e., 27 km for Shijiazhuang, 14 km for Baoding, and 7 km for Linqing. The change rate of the RGUD with distance along the urban-rural gradient is significantly higher in the large city (Shijiazhuang: 0.26 days/km) than it is in the middle- and small-scale cities (Baoding: 0.21 days/km and Linqing: 0.11 days/km), which suggests that larger cities spread heat at a faster rate than that of smaller cities. In sum, in this thesis, I (1) explored the impacts of spatial resolution on satellite-based vegetation spring phenology detection, (2) explored the impacts of the smoothing process of satellite-derived VI time series on vegetation spring phenology detection, (3) assessed the feasibility of two typical spatiotemporal data fusion algorithms on capturing the spatial and temporal variations of spring phenology at fine scales, and (4) characterized the reliability of satellite vegetation phenology detection at fine scales by investigating the urban-induced microclimate effects on winter wheat spring phenology. Overall, this thesis provided a comprehensive guideline on how to qualify uncertainties of satellite vegetation phenology detection, which improves its reliability from satellite time-series data. In addition, several practical take-home messages about how to select appropriate satellite spatial resolutions, data fusion algorithms, and optimal parameters of the VI time-series smoothing process were given to future studies on satellite vegetation phenology.