Using Penalized Linear Discriminant Analysis and Normalized Difference Indices Derived from Landsat 8 Images to Classify Fruit-tree Crops in the Aconcagua Valley, Chile

Abstract

Accurate crop type maps are critical for yield estimation and agricultural practices in modern agriculture. A new approach is proposed in this thesis to improve the crop type classification accuracy, by creating a new feature set containing new spectral indices in addition to basic bands. Two types of penalized linear discriminant analysis classifiers are adopted to do the classification, and the cross-validated classification accuracies on the two different feature sets are compared to see whether the new feature set can improve the crop identification. The result shows with new indices in the feature set the classification mean error rates were decreased substantially for both classifiers (21.6% and 25.2%). Through analyzing the coefficients retrieved from the best model, the variable importance was assessed. The coefficients are summarized by different bands and images, and the result suggest that red and shortwave infrared are the two bands highly related to the fruit-trees type identification in the study area in Aconcagua valley, Chile. Also late winter to early spring may be the best time to do crop type mapping for these crop types.