Date of Award

1984

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Earth & Environmental Systems

Abstract

This study examined the potential for utilizing Landsat spectral signatures, instead of in-situ measurements, to determine values for various physical, chemical, and biotic properties of coal mine surface features. The ten properties selected for evaluation were those that are most significantly altered by the mining activity, and which might ultimately inhibit soil-plant relationships. In-situ and spectral measurements for each of the properties were collected at 33 different mine locations. Three statistical techniques were employed to identify relationships between the two data sets. Canonical correlation analysis was used to determine the degree to which the two data sets were sensitive to the same characteristics of the mine feature properties. Multiple regression analysis was used to test the potential of the spectral data set for predicting values of the mine feature properties. Multiple discriminant analysis was also used for prediction of several of the properties which were evaluated non-metrically. The following five conclusions were derived from the study: (1) A strong relationship was found between the measurements of the two data sets, with 94.7% of the variance between them accounted for by the first canonical variate; (2) Statistically significant results were obtained using the spectral variables to predict relief, slope, vegetation type, vegetation density, parent material, surface temperature, moisture capacity, total organic carbon, and soil pH; (3) The distinct between-class differences and minimal within-class variations of the features studied resulted in only one significant spectral predictor of specific property values; (4) For minimally vegetated sites, the first principal component transformation (PC1) of the original Landsat data proved to be the best overall predictor; (5) Prediction accuracy was increased substantially when categories were substituted for specific values of the criterion variables. In general, the study successfully demonstrated that Landsat spectral data could be used to predict statistically significant in-situ measurements. However, the levels of prediction accuracy achieved using multiple regression analysis were probably insufficient for practical application. In contrast, the prediction results obtained through multiple discriminant analysis were substantially higher. Given the broad resolution of Landsat data, it may prove more applicable to use nomparametric techniques, such as multiple discriminant analysis, which facilitate generalized levels of prediction.

Share

COinS