**3.2 Models completed using visual analysis to extract the radiometric information**

Correlation analysis completed using homogeneous areas derived from visual analysis showed that spectral variables related to wetness yielded the highest Pearson coefficients. Among them, TM5, TM7, TC3, MSI and MID57 were identified as the best predictor variables, with an R > 0.68. In addition, vegetation indices were again the second group of variables that showed high correlation levels, with NDVI showing the strongest correlation

Using Remote Sensing to Estimate a Renewable Resource: Forest Residual Biomass 311

Fig. 7. The regression model selected using homogeneous areas delimited by visual analysis

The correlation analysis using the data derived from the third extraction method was only performed using the plot groups CV7 and CV6 for two reasons: (i) to directly reject in the analysis the plots with high probability of containing radiometric data related to different spectral features than the FRB data; and (ii) to identify a regression model with similar R2 values than were derived from the use of fixed 3 x 3 windows, but using a higher number of plots, making it more representative of the entire study area. In addition, only NFI-2 plots inside areas classified as pine forest in the Aragon 1:50,000 forest map were considered for this analysis. As is shown in Table 7, the radiometric variables related to wetness were again the most correlated. In addition, it was shown that the correlation coefficients increase as the homogeneity in the sample increases, independently of the performed segmentation. Focusing on these results, only the S4 segmentation in group CV6 shows higher regression coefficients than using a fixed 3x3 window without restrictions. Concretely, the maximum R value S4 was reached using the variables TM7

Nonlinear relationships between the dependent variable (FRB) and the independent variables (radiometric data) were again revealed; thus transformations were applied. Table 8 shows the regression models that were obtained using plots within group CV6. Four models were performed using a wetness radiometric variable. This type of variable was also selected initially in the only one model that used more than one variable. The R2 coefficients were situated between 0.53 and 0.59 and only one model indicated an RMSEr lower than 40% in its validation. Owing to the limited difference in terms of R2, this model with the lowest RMSEr was selected to derive FRB cartography for the study area

**3.3 Models carried out using spectral segmentation to extract the radiometric** 

and the FRB map produced applying it

**information** 

and MID57 (R>0.7).

(Figure 8).



Table 5. Pearson's coefficient of correlation (R) between spectral variables and homogeneous areas delimited by visual analysis (\* correlation is significant at the 0.05 level)

Consequently, it is possible to run estimation models using visual homogeneous areas with similar accuracy level as with fixed windows because the correlation coefficients are comparable. Therefore, the main difference in the analysis made using visual homogeneous data comes from the number of plots included, which is nearly double that of the group of percentile 4 (131 versus 68). As a result, the probability of constructing an over-fitted model is reduced (Hair et al., 1998). Therefore, this method to extract radiometric data appears to be more suitable to perform regression models for FRB estimation, as these models could be more representative of all environmental characteristics of Teruel pine forests.

Only one of the five models successfully included more than one independent variable (Table 6), owing to the high auto-correlation between them. The variable chosen in each regression attempt depends on sample divisions; but in all cases, this variable was related to wetness. The model run with the sample N3 was selected to produce a map (Figure 7), as it showed good conciliation between its R2 and its RMSEr. In addition, this model allowed direct comparison to the previous extraction method because both models use the same radiometric variable (MID57).


Table 6. Linear regression models obtained using homogeneous areas delimited by visual analysis

(Table 5). Finally, note that scatter plots comparing FRB and spectral variables also showed nonlinear relationships; therefore both independent and dependent variables were transformed using standard transformation before introducing them in the SPSS software

**Variable** *R* **Variable** *R* TM1 -0.633\* PC3 0.678\* TM2 -0.628\* TC1 -0.648\* TM3 -0.638\* TC2 0.537\* TM4 -0.330\* TC3 0.708\* TM5 -0.693\* NDVI 0.652\* TM7 -0.688\* SAVI 0.651\* PC1 -0.665\* MSI -0.708\* PC2 -0.174\* MID57 -0.692\* Table 5. Pearson's coefficient of correlation (R) between spectral variables and homogeneous

areas delimited by visual analysis (\* correlation is significant at the 0.05 level)

Consequently, it is possible to run estimation models using visual homogeneous areas with similar accuracy level as with fixed windows because the correlation coefficients are comparable. Therefore, the main difference in the analysis made using visual homogeneous data comes from the number of plots included, which is nearly double that of the group of percentile 4 (131 versus 68). As a result, the probability of constructing an over-fitted model is reduced (Hair et al., 1998). Therefore, this method to extract radiometric data appears to be more suitable to perform regression models for FRB estimation, as these models could be more representative of all environmental

Only one of the five models successfully included more than one independent variable (Table 6), owing to the high auto-correlation between them. The variable chosen in each regression attempt depends on sample divisions; but in all cases, this variable was related to wetness. The model run with the sample N3 was selected to produce a map (Figure 7), as it showed good conciliation between its R2 and its RMSEr. In addition, this model allowed direct comparison to the previous extraction method because both models use the same

*<sup>a</sup> β0 β1 β2*

N1 ln\_MID57 0.610 16.822 -3.960 - 12.207 64.95 N2 ln\_TM5 0.562 16.625 -4.541 - 7.497 41.78 N3 ln\_MID57 0.595 17.675 -4.191 - 8.839 59.48

N5 ln\_TM5 0.558 17.492 -4.838 - 11.238 56.68 Table 6. Linear regression models obtained using homogeneous areas delimited by visual

inv\_TM4 0.579 6.649 -5.909 45.999 8.079 54.32

*RMSE (ton/ha)*  *RMSEr (%)* 

for the linear regression analysis.

characteristics of Teruel pine forests.

radiometric variable (MID57).

**Model Variable** *R2*

MSI,

N4

analysis

Fig. 7. The regression model selected using homogeneous areas delimited by visual analysis and the FRB map produced applying it

#### **3.3 Models carried out using spectral segmentation to extract the radiometric information**

The correlation analysis using the data derived from the third extraction method was only performed using the plot groups CV7 and CV6 for two reasons: (i) to directly reject in the analysis the plots with high probability of containing radiometric data related to different spectral features than the FRB data; and (ii) to identify a regression model with similar R2 values than were derived from the use of fixed 3 x 3 windows, but using a higher number of plots, making it more representative of the entire study area. In addition, only NFI-2 plots inside areas classified as pine forest in the Aragon 1:50,000 forest map were considered for this analysis. As is shown in Table 7, the radiometric variables related to wetness were again the most correlated. In addition, it was shown that the correlation coefficients increase as the homogeneity in the sample increases, independently of the performed segmentation. Focusing on these results, only the S4 segmentation in group CV6 shows higher regression coefficients than using a fixed 3x3 window without restrictions. Concretely, the maximum R value S4 was reached using the variables TM7 and MID57 (R>0.7).

Nonlinear relationships between the dependent variable (FRB) and the independent variables (radiometric data) were again revealed; thus transformations were applied. Table 8 shows the regression models that were obtained using plots within group CV6. Four models were performed using a wetness radiometric variable. This type of variable was also selected initially in the only one model that used more than one variable. The R2 coefficients were situated between 0.53 and 0.59 and only one model indicated an RMSEr lower than 40% in its validation. Owing to the limited difference in terms of R2, this model with the lowest RMSEr was selected to derive FRB cartography for the study area (Figure 8).

Using Remote Sensing to Estimate a Renewable Resource: Forest Residual Biomass 313

Fig. 8. The regression model selected using segmentation and 3 x 3 pixel windows with

The accuracy assessments of every regression equation (RMSE and RMSEr) were completed using plots that showed the same homogeneity criteria as was used to run the model. However, since the selected estimation models have been applied to each one of the Landsat pixels located in forested areas in Teruel Province, the degree of success in these must also

To accomplish this, the NFI-2 plots excluded from the estimation models and their validation were considered. In order to guarantee the results, those plots that were affected by inaccuracies in their field location and/or by the radiometric response of different landscape elements located in their immediate vicinity, were removed from the validation sample. Consequently, group CV8 was used since it includes a high number of plots, which ensures that the validation results were not biased by using only the ideal

As it can be seen in figure 9, the results show few differences between the three maps. Those obtained from 3 x 3 fixed windows yieded a RMSEr of 64.26%, while spectral homogeneous forest areas had RMSEr values of 66, 71% and 65.06%, respectively. These results at pixel level can be considered tolerable for the study area considering previous experiments using similar methodologies for boreal environments less affected by heterogeneity than Mediterranean forests. Thus, Tokola et al. (1996), Tokola & Heikkila (1997), Mäkkelä & Pekkarien (2001) and Katila & Tomppo (2001) reported RMSEr to estimate forest parameters such us timber volume or total volume from about 65% to more than 100%. In this respect, it is important to emphasize that estimation error in cartography derived from satellite images decreases with an increase in the size of the area used to validate it. For example, Fazakas et al. (1999) showed a RMSEr of 66.5% at pixel level, but when using an aggregation area of 598 ha, the RMSEr was 8.7%. However, it was not possible to carry out a similar analysis in our study area because no other FRB

restrictions and the FRB map produced applying it

**3.4 FRB cartography validation** 

data were available at any scale.

be evaluated at that scale.

plots.




Table 8. Linear regression models obtained using segmentation and 3 x 3 pixel windows with restrictions

Fig. 8. The regression model selected using segmentation and 3 x 3 pixel windows with restrictions and the FRB map produced applying it
