Xie, YY; Wilson, AM (2020). Change point estimation of deciduous forest land surface phenology. REMOTE SENSING OF ENVIRONMENT, 240, 111698.
Abstract
Dramatic phenological shifts and ecosystem responses of deciduous forests to global climate change have been reported around the world. Land Surface Phenology (LSP) derived from satellite imagery is useful to estimate the phenological responses of vegetation to climate variability and inform terrestrial ecosystem models at landscape to global scales. However, there is a large (and unquantified) uncertainty in estimated phenological dates due to the relatively coarse temporal resolution of typical data and methodological limitations. To assess responses of phenology and related ecological function and services, it is essential to decrease the temporal uncertainty of estimated phenological processes. In this study, we developed a new LSP estimation method using linear change point models to determine four phenological transitions using twice-daily Moderate Resolution Imaging Spectroradiometer (MODIS) Enhanced Vegetation Index (EVI) from 2000 to 2015. We evaluated the approach using long-term phenological ground observations and compare performance of four LSP estimations generated from two data sources (i.e. 8-day and twice daily EVI time series) and two methods (i.e. double logistic and change point estimation). We found that the LSP generated from change point estimation with twice daily EVI time series had the highest accuracy (i.e. lower Root Mean Square Error (RMSE), mean bias, and Mean Absolute Error (MAE)) for both spring and fall phenology evaluated by Harvard Forest phenology observations and a large citizen science database of phenological observations from the National Phenology Network. For example, change point estimation reduced the estimation error for fall senescence date from over 40 days in the standard MODIS phenology product (version 005) to 11.5-24 days of RMSE, -2.6 to -5.8 days of mean bias, and 7.9-20.1 days of MAE. The change point methodology also enables calculation of additional metrics to describe the biophysical process of vegetation, including rates of greenup, green-down, and senescence, EVI values at each phenological transition, and the estimation uncertainties for each transition date. Our LSP estimations will improve more comprehensive investigations of landscape phenology of deciduous forest and the associated ecosystem processes at regional to global scales.
DOI:
10.1016/j.rse.2020.111698
ISSN:
0034-4257