In order to build models that relate thematic mapper (TM) imagery to field forest variables, several regression techniques, such as the ones based on the Mallows' Cp and the adjusted R2 statistics, were applied. Nevertheless, although the best created models had good fittings (R2 > 0.65) apparently supported by a clear statistical significance (p < 0.0001), later trials tested with additional plots showed that these models were, in fact, nonrobust models (models with very low-predictive capabilities). Two factors were pointed out as causes of these inconsistencies between predicted and observed values: a relatively small number of available field plots and a relatively high number of possible independent variables. Actually, different trials suggested much lower fittings for the expected really predictive models. Some restrictions of TM satellite data, such as its radiometric, spectral, and spatial limitations, together with restrictions arising from gathering and processing of field data, might have led to these poor relations. This study shows the need for guarantees stronger than the usual ones before concluding that there is a clear possibility of using satellite information to estimate forest parameters by means of regression techniques. © 1998 IEEE.
|Journal||IEEE Transactions on Geoscience and Remote Sensing|
|Publication status||Published - 1 Dec 1998|
- Multiple regression
- Remote sensing
- Thematic mapper