Feature selection and the information content of Thematic Mapper Simulator data for forest structural assessment

The information content of Thematic Mapper Simulator (TMS) data was investigated for a forested region in northern Idaho to determine the sensitivity of TMS data to forest structural characteristics (crown closure and site class). Feature selection performed via principal components analysis and a Monte Carlo simulation indicated that TMS channels 4 (0.77–0.90 μm), 7 (10.32–12.33 μm), 5 (1.53–1.73 μm), and 3 (8.63–0.69 μm) were the four optimal channels for forest structural analysis. These four channels utilized the full spectral capability of the Thematic Mapper, representing wavelengths from the visible, the near-infrared (IR), the mid-IR, and the thermal portions of the electromagnetic spectrum. As the number of channels supplied to the Monte Carlo feature selection routine increased, classification accuracy increased. The information content of the TMS data was analyzed by performing supervised maximum likelihood classifications on three data sets: 1) 7-channel 30-m 8-bit data, 2) the 4-optimal-channel 30-m 8-bit data, and 3) TMS data degraded to Landsat multispectral scanner (MSS)specifications, 3-channel 60-m 6-bit data. The greatest sensitivity to forest structural parameters, which included crown closure, site class, and succession within clearcuts, was obtained from the 7-channel TMS data, the 4-optimal-channel TMS data, and the simulated MSS data, respectively. The increased number of spectral hands was largely responsible for the increased accuracy of the TMS data over the simulated MSS data. The improved spatial resolution of the TMS data did not improve classification performance. Variance within the TMS scene was largely due to the structural characteristics of the forest canopy.