In this paper, we deal with the problem of extracting meaningful textural features leading to good segmentations on satellite images of natural environments. Standard texture features using gray level co-occurrence matrices have been widely applied on remote sensed images but they impose limitations (due to finite window sizes) as poor spatial localization. We have generalized the definition of texture features using a multiscale framework, in order to take advantage of multiscale properties of natural images. The new definition improves spatial localization and the relevance of the parameters. We then investigate the dependencies among different features for classification purposes. An unsupervised scheme of classification was performed on different satellite infrared images. We see that natural, chaotic images should be treated with a different methodology.