For many ecological applications of remote sensing, traditional multispectral data with moderate spatial and spectral resolution have to be used. Typical examples are land-use change or deforestation assessments. The study sites are frequently too large and the timespan covered too long assumes the availability of modern datasets such as very high resolution or hyperspectral data. However, in traditional datasets such as Landsat data, separability of the relevant classes is limited. A promising approach is to describe the landscape context pixels that are integrated. For this purpose, multiscale context features are computed. Then, spectral-spatial classification is employed. However, such approaches require sophisticated processing techniques. This study exemplifies these issues by designing an entire framework for exploiting context features. The framework uses kernel-based classifiers which are unified by a multiple classifier system and further improved by conditional random fields. Accuracy on three scenarios is raised between 19.0%pts and 26.6%pts. Although the framework is designed, focusing an application in Chile, it is generally enough to be applied to similar scenarios.