Application of spectral features' ratios for improving classification in partially calibrated hyperspectral imagery: a case study of Mediterranean vegetation species separation
Ronit Rud, Maxim Shoshany, Victor Alchanatis, Yafit Cohen
Introduction
Limitations on radiometric calibration of spectral imagery highly degrade their capabilities to efficiently detect objects in the imaging field and determining their properties. This is especially true for real time spectral imaging which on one hand have inherent constraints regarding exposure time and data transfer rates in acquiring high dimensional hyperspectral cubes, and on the other hand, are required to operate in non-ideal conditions: partially cloud overcast, quick changes in the source radiation field etc.
Image Radiometric calibration in most applications treats the scene plane as homogenously illuminated (scene plane calibration): not accounting for effects from flux density variations on the surfaces of objects protruding from this plane (e.g., Shoshany, 1992, 1993). This obviously influence objects' separability and causes inaccurate and false classifications. Calculation of the real flux densities would require information regarding the objects' 3-D structure, which would not be available in most real-time imaging applications. Rationing techniques, which are very simple to implement in real-time sensing applications, may help in partially overcoming this obstacle. Analysis of the spectral properties is then required for discovering the optimal spectral band ranges as well as their features to be applied. This reported study suggests improvements in using scene plane calibrated images by addressing these two integrated issues: 1) determining the spectral features (e.g.: derivatives, absorption depth, absorption symmetry, etc.); and 2) their spectral definition bounds. Clustering in the new feature space is then hypothesized to improve the classification objects from real-time imagery.
This case study consists of eight Mediterranean vegetation species characterized by high spatio-temporal diversity in their spectral reflectance due to illumination effects as well as seasonal and habitat conditions effects. The discrimination of species is a highly challenging task, especially regarding real-time imaging. Spraying machines for site specific spraying which need to adjust spraying (position and quantities) to each plant are an example for real-time imaging applications. Success in the discrimination of natural species is of relevance not only for real-time applications but rather for long term ecological monitoring under different climatic conditions.
The aims of this study are: the development of techniques facilitating the discrimination of objects from spectral imaging calibrated only for generalized scene radiometry, and their testing in separation of species from images of natural scenes.
Figure 4: Reflectance of Mediterranean species (a), and their 1st derivative (b): green and red vertical lines represent the ranges’ boundaries – green peak (F1) and read absorption (F2), Max. & Min. indicate location of maximum and minimum derivatives’ values for F1 and F2
Figure 5: Continuum removal results for the two spectral features: inversed green peak (a) and red absorption (b).
|
Hyper Spectral 
