Using active learning to adapt remote sensing image classifiers

The validity of training samples collected in field campaigns is crucial for the success of land use classification models. However, such samples often suffer from a sample selection bias and do not represent the variability of spectra that can be encountered in the entire image. Therefore, to maximize classification performance, one must perform adaptation of the first model to the new data distribution. In this paper, we propose to perform adaptation by sampling new training examples in unknown areas of the image. Our goal is to select these pixels in an intelligent fashion that minimizes their number and maximizes their information content. Two strategies based on uncertainty and clustering of the data space are considered to perform active selection. Experiments on urban and agricultural images show the great potential of the proposed strategy to perform model adaptation.

Research highlights
► Changes in illumination or geometry make difficult to transfer classification models.
► Active learning is proposed to adapt classification models to new similar images.
► A criterion based on model uncertainty is proposed to adapt the known distribution.
► A criterion based on clustering is proposed to find new, unknown, classes.
► In four high resolution images, active sampling ensures fast convergence and optimal adaptation.