Accuracy of Forest Stem Volume Estimation by TM/Landsat Imagery with Different Geometric and Atmospheric Correction Methods
Elias Fernando Berra, Denise Cybis Fontana, Rudiney Soares Pereira

Abstract
This study aims to evaluate the accuracy in the indirect estimation of stem volume of Pinus from TM/Landsat 5 images, which were processed with different geometric and atmospheric correction methods. Regressions were used to estimate the stem volume (m³/ha), where the independent variable was the value of NDVI (Normalized Difference Vegetation Index) related to the forest sampling unit measured in the field. The surface spectral reflectance used in the NDVI calculation were obtained by four different methods: 1) geometric correction with nearest-neighbor (NN) resampling + atmospheric correction using dark object subtraction (DOS), 2) NN resampling + atmospheric correction using MODTRAN (Moderate Resolution Transmittance), 3) geometric correction with bilinear resampling + DOS, and 4) bilinear resampling + MODTRAN. The reliability of the estimates was measured by means of bias (Bias) and standard error (RMSE). Among the atmospheric correction methods, the errors were higher with DOS. Regarding the geometric correction methods, the RMSE and Bias were higher with the NN resampling. Thus, the combination of DOS + NN had the highest RMSE (62.3%) and Bias (15.9%). The best estimate was the combination of bilinear resampling + MODTRAN, with RMSE of 56.5% and Bias of 13.3%. Therefore, it was verified that different methods of geometric and atmospheric correction should be tested to improve the estimates of forest biophysical variables.

Full Text: PDF