Spectral-Temporal Classification Using Vegetation Phenology

Abstract

In this paper we describe a multitemporal classification procedure for crops in Landsat scenes. The method involves the creation of crop signatures which characterize multispectral observations as functions of phenological growth states. The phenological signature models spectral reflectance explicitly as a function of crop maturity rather than a function of date. This means that instead of stacking spectral vectors of one observation on another, as is usually done for multitemporal data, for each possible crop category a correspondence of time to growth state is established which minimizes the smallest difference between the given multispectral multitemporal vector and the category mean vector indexed by growth state. The results of applying it to winter wheat show that the method is capable of discrimination with about the same degree of accuracy as more traditional multitemporal classifiers. It shows some potential to label degree of maturity of the crop without crop condition information in the training set.