Ensemble of Support Vector Machines for Change Detection in Remotely-Sensed Land Cover Images

Change detection plays a vital role in satellite surveillance. With increase in the availability of multi-temporal and multi-spectral images, change detection is considered to be the biggest challenge in the field of remote-sensing. In this work, a semi-supervised change detection technique is proposed for the classification of multi-temporal remotely-sensed images. This technique uses an ensemble of multiple Support Vector Machines (SVM) to model the spatial regularity between pixels of the multi-temporal difference image (DI) to classify them as change or no-change. SVM’s learning performance is significantly influenced by the selected kernel function. This work presents an ensemble-SVM approach that avoids the problem of choosing an optimum kernel. Several SVMs, each with a different kernel function, have thus been ensembled to produce a model that can predict whether a pixel in the DI is to be classified as change or no-change. Compared against single kernel SVMs, this work shows a significant increase in the precision and robustness of the ensemble SVM technique.