- Identification of Optimal Field Spectral Measurements Wavebands for Discriminating among Spatial Features in support of Mapping Using Hyper-spectral Imagery
Identification of Optimal Field Spectral Measurements Wavebands for Discriminating among Spatial Features in support of Mapping Using Hyper-spectral Imagery
Full Paper Review
Mwitwa Chilufya,University of KwaZulu-Natal, South Africa, email@example.com
Mulemwa Akombelwa, University of KwaZulu-Natal, South Africa, firstname.lastname@example.org
Derek Stretch, University of KwaZulu-Natal, South Africa , email@example.com
Field spectral measurements serve as ground truthing data necessary for classification of hyper-spectral imagery. For this purpose, optimal wavelengths capable of discriminating all or the majority of features sampled in the field need to be identified prior to hyper-spectral imagery classification. This paper discusses the identification of optimal wavebands from field spectral measurements for the purpose of mapping spatial features using hyper-spectral imagery. Spectral Measurements of several vegetation assemblages in the Mfabeni Wetland of the Isimangaliso Wetland Park were collected at three different sites using an Analytical Spectral Devices (ASD) spectralradiometer, then pre-processed to a format suitable for processing using Random Forest (RF), an open-source machine learning software package that runs on the statistical package R.
The pre-processed data were normalized to Hyperion imagery spatial resolution, and then used to develop an optimal model that identifies the most suitable wavebands for discriminating pixels representing the vegetation assemblages collected at one of the three sites. The resultant model was then used to predict pixels on the other two sites with similar spectral characteristics. Results show that RF can be used to develop a model for discriminating optimal wavebands for mapping vegetation assemblages at a pixel level and that the developed model can be used to predict pixels with similar spectral characteristics in an independent dataset of spectral measurements. This approach can, by implication, be used for any given data set of spectral measurements, thus allowing for mapping of any combination of spatial features using hyper-spectral imagery.