Plant species discrimination using emissive thermal infrared imaging spectroscopy

Discrimination of plant species in the optical reflective domain is somewhat limited by the similarity of their reflectance spectra. Spectral characteristics in the visible to shortwave infrared (VSWIR) consist of combination bands and overtones of primary absorption bands, situated in the Thermal Infrared (TIR) region and therefore resulting in broad spectral features. TIR spectroscopy is assumed to have a large potential for providing complementary information to VSWIR spectroscopy. So far, in the TIR, plants were often considered featureless. Recently and following advances in sensor technology, plant species were discriminated based on specific emissivity signatures by Ullah et al. (2012) using directional-hemispherical reflectance (DHR) measurements in the laboratory. Here we examine if an accurate discrimination of plant species is equally possible using emissive thermal infrared imaging spectroscopy, an explicit spatial technique that is faster and more flexible than non-imaging measurements.

Hyperspectral thermal infrared images were acquired in the 7.8⿿11.56 μm range at 40 nm spectral resolution (@10 μm) using a TIR imaging spectrometer (Telops HyperCam-LW) on seven plants each, of eight different species. The images were radiometrically calibrated and subjected to temperature and emissivity separation using a spectral smoothness approach. First, retrieved emissivity spectra were compared to laboratory reference spectra and then subjected to species discrimination using a random forest classifier. Second, classification results obtained with emissivity spectra were compared to those obtained with VSWIR reflectance spectra that had been acquired from the same leaf samples.

In general, the mean emissivity spectra measured by the TIR imaging spectrometer showed very good agreement with the reference spectra (average Nash-Sutcliffe-Efficiency Index = 0.64). In species discrimination, the resulting accuracies for emissivity spectra are highly dependent on the signal-to-noise ratio (SNR). At high SNR, the TIR data (Overall Accuracy (OAA) = 92.26%) outperformed the VSWIR data (OAA = 80.28%).

This study demonstrates that TIR imaging spectroscopy allows for fast and spatial measurements of spectral plant emissivity with accuracies comparable to laboratory measurement. This innovative technique offers a valuable addition to VSWIR spectroscopy as it provides complimentary information for plant species discrimination.