Detecting the nature of change in an urban environment: A comparison of machine learning algorithms

Abstract

The performance of difference machine learning algorithms for detecting nature of change was compared. To alleviate the problem of obtaining enough training data, simulated training data were generated from single-date images. A one-pass classification with four machine learning algorithms, namely, Multi-Layer Perceptrons (MLP), Learning Vector Quantization (LVQ), Decision Tree Classifiers (DTC), and the Maximum- Likelihood Classifier (MLC), were tested. Recognition rates, ease of use, and degree of automation of the four algorithms were assessed. The results showed that the incorporation of cross-combined simulated training data enhanced the detection of nature of change. Compared to conventional post-classification comparison methods, LVQ and DTC did better in terms of overall accuracy. In terms of average accuracy of the change classes, LVQ was the best performer. DTC was the easiest to use and the most robust in training. MLP procedures were the most difficult to replicate.