Value of feature reduction for crop differentiation using multi-temporal imagery, machine learning, and object-based image analysis.

Abstract

This study examined the value of automated and manual feature selection, when applied to machine learning and object-based image analysis (OBIA), for the differentiation of crops in a Mediterranean climate. Five Landsat8 images covering the phenological stages of seven major crops types in the study area (Cape Winelands, South Africa) were acquired and processed. A statistical image fusion technique was used to enhance the spatial resolution of the imagery. The pan-sharpened imagery was used to produce a range of spectral features, textural measures, indices and colour transformations, after which it was segmented using the multi-resolution (MRS) algorithm. The entire set of 205 features (41 per image capture date) was then subjected to different feature selection and reduction methods. The feature selection and reduction methods included manual feature removal (i.e. grouping into semantic themes), filter methods (such as classification and regression trees (CART) and random forest (RF)), and statistical principal components analysis (PCA). The experiments were carried out in two scenarios, namely 1) on all input images in combination; and 2) on each individual image date. The feature subsets were used as input to decision trees (DTs), k-nearest neighbour (k-NN), support vector machine (SVM), and random forest (RF) machine learning classifiers. In order to assess the value of each feature reduction method (comprising feature reduction and selection techniques), overall accuracy, kappa coefficient and McNemar’s test were employed to assess classification accuracy and compare the results. The results show that feature selection was able to improve the overall crop identification accuracy for the DT, k-NN, and RF classifiers, but was unable to do so for SVM. SVM scored the highest overall accuracy and kappa coefficient, even without applying feature reduction or selection.