**2.5 FRB mapping and validation method**

306 Remote Sensing of Biomass – Principles and Applications

Fig. 4. (a) Homogeneous area delimited using high-resolution aerial photographs; (b) Core pixels of the homogeneous area used to extract spectral data from the Landsat TM image

The main problems in constructing predictive models related to the extraction of radiometric data using high-resolution aerial photographs of delimited forest area are: (i) human errors in the establishment of limits during the visual image interpretation process, (ii) assumption of a FRB constant value for the entire homogeneous area based on a point

Image segmentation can be applied to delimit homogeneous spectral features as a reference for spectral data extraction (Hall et al., 2006; Mäkelä & Pekkarinen, 2001; Pekkarinen, 2002). In this study, the segmentation algorithm RGB clustering incorporated in the software Erdas Imagine was applied. The essential parameters that control the results in this method are the RGB composition (which is used to apply the segmentation), the stretch applied to each band in the composition, and the number of bins into which each band is divided. Four different segmentations were performed, modifying the stretch (2 or 4 standard deviations) and the bins (7-6-6 and 4-3-3) using a RGB composition TM7-TM4-TM3 to model the heterogeneity of the analyzed forest (Table 2). In addition, in order to remove pixels within the immediate vicinity that were not attributable to the IFN-2 plot, the obtained image segments containing field plots were then intersected with 3 x 3 pixel windows, resulting in a mask of homogeneous pixels. Thus, the mean pixel value for each image band was then computed from the pixels that belong to the same spectral category as that of the central

S1 2 7 6 6 S2 4 7 6 6 S3 2 4 3 3 S4 4 4 3 3

Table 2. Parameters considered in the four different segmentations performed using RGB

**Number of bins Red Green Blue** 

location that is not representative of larger forest stands (Pekkarinen, 2002).

**2.4.3 Spectral segmentation** 

pixel of the NFI-2 plot (Figure 5).

composition TM7-TM4-TM3

**Segmentation Number of Std. dev. in** 

**each band** 

We used Pearson's coefficient to explore the feasibility of building accurate and representative predictive models using the plot groups derived from the three extraction methods. After this selection, each of the three groups was divided into two samples: 80% of the sample was used to carry out the predictive model and the remaining 20% was used to validate it. This sample division in each group was made randomly to guarantee the execution of the estimation equations and the validation processes. Moreover, to assess the robustness of the models, the sample division was done five times, completing the respective estimation model and its validation each time.

The multiple linear regression model was performed, using the option "stepwise" to include only the significant variables (p < 0.05). In addition, performance was verified for all of the principles assumed for this type of regression at the model and variable level. To evaluate the performance of models, the coefficient of determination (R2) was used, and the Root Mean Square Error (RMSE) and the relative RMSE (RMSEr) were calculated using 20% of the sample previously reserved for the validation. Finally, the best model conducted with each of the extraction methods was applied to the 1994 Landsat data in order to obtain the FRB estimation cartography. The three derived cartographies were validated using plots not included in the groups considered in the regression models, and the RMSE and RMSEr were calculated to evaluate the results.

### **2.6 Model application and inventory**

The results obtained with the June 1994 Landsat image (selected on the basis of its temporal coincidence with NFI-2 fieldwork) at the regression model level and in the cartography validation (in terms of R2 and RMSE and RMSEr, respectively) were analyzed. Then, the most suitable estimation model was selected for application to the July 2008 Landsat image. As a result, current information was obtained about the potential quantity of this energy resource and its spatial distribution in the study area.
