Forest change detection applying Landsat Thematic Mapper difference features: a comparison of different classifiers in boreal forest conditions

This article addresses the problem of detecting forest changes via multitemporal Landsat Thematic Mapper (TM) imagery. A stand-level classification approach is selected, where, for each stand, a total of 35 statistical differences is extracted from Landsat TM images. Three forest stand-change classes are considered: 1) no change, 2) moderate change, and 3) considerable change. The classification results are reported by using the following classifiers: K-nearest-neighbor, Maximum Likelihood classifier with Gaussian- and Kernel-based class probability density estimation, Classification and Regression Trees, Multilayer Perceptron (MLP) early stop committee, and the MLP with weight decay training. Two Bayesian learning methods for MLP are also used: The Evidence Framework of MacKay, and Hybrid Monte Carlo (HMC) method following Neal. The best overall correct classification result (88.1%) is obtained by MLP trained with HMC and the Automatic Relevance Detection approach, but the variation of the performance of the classifiers is rather small.