Pattern recognition of landsat data based upon temporal trend analysis

Abstract

The problem of deciding whether or not a classification of a Landsat agricultural scene is acceptable when no ground truth is available was addressed. The approach taken was to examine temporal trends of the Landsat mean vectors of crops. A procedure for agricultural crop classification was developed using a time series (multitemporal) of Landsat mean vectors for selected agricultural fields in Montana and Kansas for which ground truth was known. This procedure using the temporal trend of mean vectors (the temporal trend procedure) was then applied to the individual Landsat pixels in more than one hundred multitemporal data sets collected throughout the wheat growing regions of the United States. The resulting classifications have compared favorably to ground truth estimates for proportion of wheat in those cases where ground truth was available. This temporal trend procedure has been found to give estimates of the wheat proportion that are comparable to the best results obtained using maximum likelihood classification with photointerpreter defined training fields. This classification scheme utilizing a temporal trend procedure is referred to as the “Delta Classifier”. It is currently being used as an independent, end-of-the-growing-season check on the reasonableness of maximum likelihood results in a quasi-operational Large Scale experiment (MacDonald et al., 1975).