Classifying fruit-tree crops by Landsat-8 time series

Abstract

Landsat-8 time series were used to classify major crops types in Maipo and Aconcagua valleys, central Chile. In the former valley four fruit-tree crops were classified applying different machine learning techniques on feature sets comprising typical index-based temporal profiles, like those using the normalized difference vegetation index, and the complete spectral resolution of the time series. In the latter valley six fruit-tree crops were classified only by LDA (linear discriminant analysis), found the best performing classifier for the Maipo Valley. LDA was applied on the complete spectral resolution of the time series and on a feature set adding all possible NDIs (normalized difference indices) that can be constructed from the time series. Regardless of the feature set used good MERs (misclassification error rates) were found (≤ 0.21) for the Maipo's crops, but they were reduced by 4 and 13 percentage points, depending on the classifier and the training sample size used, when using the complete spectral resolution of the time series. We further explored these findings in the Aconcagua Valley, where MERs were reduced from 0.13 to 0.1 when the NDI-based feature set was used. In both study cases, the most predictive bands belonged to the first image dates of the time series, corresponding to the crops' greenup stage, and they were placed not only on the typical greenness spectral region but also on the shortwave infrared region.