干旱区生态环境知识资源中心

Vegetation phenology is commonly studied using time series of multi-spectral vegetation indices derived from satellite imagery. Differences in reflectance among land-cover and/or plant functional types are obscured by sub-pixel mixing, and so phenological analyses have typically sought to maximize the compositional purity of input satellite data by increasing spatial resolution. We present an alternative method to mitigate this ’mixed-pixel problem’ and extract the phenological behavior of individual land-cover types inferentially, by inverting the linear mixture model traditionally used for sub-pixel land-cover mapping. Parameterized using genetic algorithms, the method takes advantage of the discriminating capacity of calibrated surface reflectance measurements in red, near infrared, and short-wave infrared wavelengths, as well as the Normalized Difference Vegetation Index (NDVI) and the Normalized Difference Water Index. In simulation, the unmixing procedure reproduced the reflectances and phenological signals of grass, crop, and deciduous forests with high fidelity (RMSE < 0.007 NDVI); and in empirical tests, the algorithm extracted the phenological characteristics of evergreen trees and seasonal grasses in a semi-arid savannah. The approach shows potential for a wide range of ecological applications, including detection of differential responses to climate, soil, or other factors among vegetation types.