Predicting forest stand attributes using multispectral and hyperspectral satellite data in southwestern Ohio

Background/Question/Methods

In this study, data from Landsat TM, SPOT 5, and EO-1 hyperspectral imager (Hyperion) were evaluated for their suitability to predict species richness and basal area and eventually mapping five forest types in southwestern Ohio. Forest types studied ranged from remnants of old-growth beech-maple forests to early successional juniper forests. Field data were gathered from 0.1 ha plots located in representative patches of the five forest types. Within each plot, forest stand data including composition by species, stem density, and basal area were obtained and plot importance values calculated. Classification and ordination techniques were used to identify species assemblages. Spectral bands from each sensor were corrected for atmospheric effects and converted to top-of-the atmosphere reflectances. Reflectance data corresponding to the 28 plots used in this study were correlated to basal area, species richness and evenness. Results/Conclusions

For Landsat TM, band 2 (acquired in the green region) was found to have the highest correlation with species richness (r = 0.559, P < 0.01), and Basal area was best correlated with TM band 3 (r = 0.459, P < 0.01). None of the Landsat bands and derived vegetation indices were correlated with species evenness. Landsat images obtained earlier in the summer (May) produced better correlations with various stand parameters compared to those obtained in late summer (August) and those acquired in the leaf-off period (April). Many Hyperion bands located between the blue and the red regions of the electromagnetic spectrum were significantly correlated with species evenness (r = 0.382-0.514, P < 0.01). A majority of Hyperion bands acquired in the NIR region were significantly correlated with basal area (r = 0.464 – 0.547, P < 0.01). Landsat TM and hyperion reflectances corresponding to groups of plots, identified as distinct communities types using classification and ordination of vegetation data, were extracted from the images and analyzed in order to determine whether these groups were spectrally separable using the Jeffries-Matusita distance. Those bands found best at separating the different community types for each sensor were then used in image classifications and the resulting maps compared. Landsat TM bands 2, 4, and 5 were found to produce the best separability between the vegetation types identified through ecological classifications.