LARS Tech Report Number
030573
Abstract
Increasing demands on the forest resource will necessitate increasingly more intensive management in the future. In order to achieve this goal, reliable and timely information over large geographic areas will be required. Remote sensing techniques offer much potential for the procurement of such information.
This research, then, was pointed toward study of that potential. Four objectives were established as follows: 1) to determine the optimum number of the available twelve multispectral scanner (MSS) wavelength bands to use for forest cover mapping with automatic data processing (ADP) techniques; 2) to determine the current capability to map basic forest cover using MSS data and ADP techniques; 3) to determine the relative utility, to forest cover mapping, of the four spectral regions available in the twelve-channel MSS data (i.e. visible, and near, middle and thermal infrared); and 4) to compare the accuracy of digitized color infrared photography with that of MSS data for forest cover mapping using ADP techniques.
In attaining the first objective, statistics defining the six cover type classes of interest (deciduous forest, coniferous forest, water, forage, corn, and soybeans) were calculated and used by the computer as a basis for the selection of "best" wavelength band combinations ranging in size from one through ten wavelength bands each. With the spectral information contained in each of these combinations, and with all twelve channels, the entire test area was classified into the six defined classes, using the LARSYS programs. Tests of the computer's performance indicated that the use of five wavelength bands would fulfill the dual requirements of adequate accuracy and moderate computer time.
In fulfilling the second objective, the automatically selected "best" combination of five channels (one each from the green and red visible wavelengths, and the near, middle and thermal infrared wavelengths) produced classification accuracies in excess of 90 percent for deciduous and coniferous forest. When these two classes were grouped, the accuracy for combined forest was in excess of 95 percent. The use of all twelve channels caused only a slight increase in overall accuracy.
In satisfying the third objective, the LARSYS feature selection processor was allowed to consider wavelength bands constituting only various subsets of the four spectral regions. On this basis, it selected a number of five-channel combinations. Classifications performed by these various channel combinations indicate that the visible wavelengths are sufficiently accurate for classifying combined forest, but inadequate for differentiating between deciduous and coniferous forest. The infrared channels separated the two forest classes with reasonable accuracy, but allowed confusion between forest and the agricultural classes. The deletion of either the near or the middle infrared individually, did not reduce accuracies, but, when both were deleted, accuracies dropped drastically. The deletion of the thermal infrared had little effect on forest cover mapping but did allow considerable confusion among the agricultural cover types. These results indicate that the thermal infrared is desirable, but not necessary, for basic forest cover mapping, and that accurate classification of deciduous and coniferous forest cover can be achieved with the visible plus either the near or middle infrared spectral regions.
To meet the fourth objective, small-scale color infrared photography, acquired the same day over the test site, was color separated, digitized in three wavelength bands and, automatically classified. In general, the digitized photography was inadequate for automatic forest cover mapping and compared poorly to the MSS data results when similar wavelength bands were used. These results were apparently caused by the narrower dynamic range, poorer spectral resolution, and uneven illumination (due to vignetting and the anti-solar point) of the photographic data.
Date of this Version
January 1973