Classification

This case study aims to show how homonim can assist in improving classifier performance. It uses images and ground truth from a study that mapped canopy cover of spekboom (a species of succulent shrub) in the Little Karoo, South Africa.

The data consists of 4 raw and corrected NGI images with 50 cm spatial resolution and 4 spectral bands (red, green, blue and near-infrared) . These images are a sub-set of the mosaics described in the aerial mosaic correction case study. Ground truth is made up of 161 polygons with labels for 3 vegetation classes:

Class	Description
Spekboom	A species of succulent shrub.
Tree	Woody trees.
Background	Other vegetation, bare ground etc.

Some example polygons are shown below, overlaid on an aerial image.

Class polygons

Correction

The corrected images were produced with a gain model, a 1 x 1 pixel kernel, and a MODIS NBAR reference image. Some detail on these choices is given in the aerial mosaic correction study. Using the blue and NIR pixel values as features, class densities (Gaussian kernel density estimates) are shown before and after correction below.

Class densities

The classes appear more compact, and likely better separated after correction.

Evaluation

A basic classifier was evaluated on the source and corrected imagery to quantify the effect of surface reflectance correction. Evaluation used the red, green, blue and NIR pixel values as features in a naive Bayes classifier, and a 10-fold cross-validation for testing. Accuracy increased from 58% to 67% demonstrating an improvement in the descriptive power of the image.

Normalised confusion matrix, accuracy, and AUC values for source and corrected images are tabulated below.

	Confusion matrix	Accuracy	AUC
Source	Background Spekboom Tree Background 0.43 0.13 0.43 Spekboom 0.09 0.62 0.29 Tree 0.07 0.23 0.70	58.41 %	0.73
Corrected	Background Spekboom Tree Background 0.56 0.17 0.28 Spekboom 0.07 0.72 0.21 Tree 0.04 0.21 0.74	67.47 %	0.81