Farid Qamar

Using ground-based, remote hyperspectral images from 0.4–1.0 micron in ∼850 spectral channels—acquired with the Urban Observatory facility in New York City—we evaluate the use of one-dimensional Convolutional Neural Networks (CNNs) for pixel-level classification and segmentation of built and natural materials in urban environments. We find that a multi-class model trained on hand-labeled pixels containing Sky, Clouds, Vegetation, Water, Building facades, Windows, Roads, Cars, and Metal structures yields an accuracy of 90–97% for three different scenes. We assess the transferability of this model by training on one scene and testing to another with significantly different illumination conditions and/or different content. This results in a significant (∼45%) decrease in the model precision and recall as does training on all scenes at once and testing on the individual scenes. These results suggest that while CNNs are powerful tools for pixel-level classification of very high-resolution spectral data of urban environments, retraining between scenes may be necessary. Furthermore, we test the dependence of the model on several instrument- and data-specific parameters including reduced spectral resolution (down to 15 spectral channels) and number of available training instances. The results are strongly class-dependent; however, we find that the classification of natural materials is particularly robust, especially the Vegetation class with a precision and recall >94% for all scenes and model transfers and >90% with only a single training instance.