Creating Models of Custom Image Classification Workflows Using Softmax Regression and Support Vector Machine

Abstract

This study demonstrates how complex image classification workflows can be built using a visual modelling tool. Models facilitate the comparison of different classifiers while allowing an analyst to experiment with different input features. The models include custom workflow steps for preparing input and training data, training the classifier, classifying images and evaluating the results. The example models presented here were used to classify Sentinel-2 imagery of eastern Texas, USA, into five land-use categories that consisted primarily of vegetation. Separate models were created for Softmax Regression and Support Vector Machine (SVM) classification, each using Sentinel-2 spectral bands and again with an additional entropy texture image as input. The results showed that SVM performed better than Softmax Regression and that the selected texture measure did not improve classification results. A discussion is provided of how the models could be extended further to provide different analysis options.