Study area

The study area comprises temperate forests of the municipality of Pueblo Nuevo, specifically in the El Brillante, La Ciudad, La Victoria and La Campana ejidos, in the state of Durango, Mexico (Figure 1). It is located in the physiographic province of the Western Sierra Madre and in the subprovince known as the Grand Plateau and the Durango Canyons. Its altitudes vary between 1 400 and 2 600 masl. The climate is semi-humid temperate or semi-cold, becoming temperate or semi-dry on the eastern side of the Sierra. The mean annual temperature ranges between 20 °C and 22 °C, and the mean annual precipitation, between 800 msm and 1 200 mm. Because of its geographical location, it includes various vegetation conditions, among them pure oak and pine-oak forests (^{Ucodefo 6, 1997}).

Figure 1. Geographical location. 

Field work

The carbon reservoirs were quantified based on the dasometric data of thirty-nine 50 x 50 m quadrangular sites belonging to the Network of Permanent Forest Research Plots established by the Universidad Juárez del Estado de Durango and the Instituto Tecnológico de El Salto (^{Corral et al., 2009}). The biomass of each sampling site was estimated using allometric equations previously generated for the main species (biomass production) of the state of Durango, which were applied to the dasometric data collected at the sampling sites. The biomass results were converted to stored amounts of carbon by multiplying the biomass by the specific total carbon factor for each species. The values of these percentages range between 0.43 and 0.51. The models were validated using 10 permanent sampling sites established in the La Campana ejido because the ecological and silvicultural conditions of this area resemble those of the El Brillante, La Ciudad and La Victoria ejidos.

The biomass and total carbon contents in each sampling site were determined by adding the amounts for both factors per tree. Furthermore, the results of these variables from the 39 sites were extrapolated to tons per hectare in order to facilitate the generation and construction of the models so as to predict their estimation through remote perception.

Landsat TM (Thematic Mapper) spectral data

The satellite image was based on the United States Geological Service (^{USGS, 2013}) and was taken on June 19, 201, with a spatial resolution of 30 m. It has been orthorectified, i.e. it represents the objects in the site at their actual coordinates. This sensor operates with seven spectral bands especially selected to monitor the vegetation, with the exception of band 7, which was added for geological applications (Table 1).

Table 1 Characteristics of the bands and main applications of the Landsat TM. 

Bi = Spectral bands; CIR = Close infrared; MIR = Medium infrared; TIR = Thermal infrared; μm = Microns.

The spectral data were taken from the central UTM coordinates of the sampling sites, which were identified by the pixels of the images. The spectral data consisted in the reflectances obtained from four individual bands of the image and four mathematical transformations applied to the reflectance: 1) Normalized difference vegetation index (NDVI) = CIR-R)/(CIR+R) (^{Rouse et al., 1974}); 2) NDVI41, consisting of the medium infrared (MIR) and green (G) spectral bands; calculated as NDVI = (MIR-G)/(MIR+G); 3) NDVI42, calculated as NDVI42 = (MIR-R)/(MIR+R); it is made up of the medium infrared and red spectral bands; 4) NDVI43, calculated as NDVI43 = (MIR-CIR)/ (MIR+CIR); this index is known as hydric stress (^{Rock et al., 1986}), and consists of the medium infrared and close infrared spectral bands. These indices were built in order to highlight certain characteristics of the vegetation, mainly in terms of chlorophyll, the cell structure, and the water content of the vegetation (^{Chuvieco, 1998}; ^{Soria and Granados, 2005}; ^{Aguirre et al., 2009}); they are most sensitive at a given portion of the electromagnetic spectrum and can be best detected in specific bands (^{Cohen and Goward, 2004}).

Relationship between field data and spectral data

An analysis of the correlation between the forest variables of interest and the spectral data was performed in order to determine the behavior and relationship between them. The stepwise statistical procedure was applied in order to select a multiple linear regression model by which to estimate the biomass (Mg ha^-1) and carbon (Mg ha^-1) with a minimum of variables. The coefficient or determination (R²), the root mean square error (RMSE), and a significance level below 0.05 for each variable were used as statistical indicators of the goodness of fit. Once the best regression models were developed, the biomass and the carbon were estimated for the sampling sites corresponding with the spectral data, and, based on them, biomass maps and aboveground arboreal carbon maps were generated for the studied locations.

Validation

With the support of the methodology described above, the dasometric data and central UTM coordinates of 10 sampling sites other than the 39 used for building the models -located within the La Campana ejido- were processed. The results obtained through the application of the predicted regression models were compared to the observed field data, estimated by means of allometric equations and carbon factors.

Relationship between the forest variables and the spectral data

The forest variables biomass and carbon showed a negative correlation with the spectral data (Table 2). According to ^{Hall et al. (2006)} and ^{Aguirre et al. (2009)}, this is accounted for by the decreased albedo in areas with dense, closed vegetation. Although little correlation was found between the biomass and the carbon with the reflectances of bands 1, 2, 4, and the NDVI42 vegetation index, the NDVI43 vegetation index had the highest correlation, with a value of -0.62 for biomass and -0.63 for carbon. ^{Aguirre et al. (2009)} recorded a correlation value of -0.80 between the NDVI43 and the aboveground arboreal carbon. The results obtained confirmed the findings of ^{Gong et al. (2003)} and ^{Aguirre et al. (2009)}: that the density of conifer forests (with needle leaves) is better accounted for by moisture sensitive indices than by chlorophyll sensitive indices. The negative sign indicates that the forest density is inversely proportional to the hydric stress of the vegetation (NDVI43). Therefore, the NDVI has good correlations with the forest variables of interest, due to the reflectivity caused by the vegetation’s chlorophyll. As observed by ^{Chuvieco (1998)} and ^{Soria and Granados (2005)}, when the vegetation suffers from any type of stress (pests, drought, etc.), its reflectance in NDVI is lower, and therefore, an area with healthy vegetation will have a higher contrast between the bands that make up this index (close infrared and red).

Table 2 Coefficients of correlation between the forest variables and the spectral data. 

B = Biomass (Mg ha^-1); C = Carbon (Mg ha^-1); Bi = Reflectance by spectral band; NDVI = Normalized difference vegetation index.

Selected models

Vegetation indices based on the reflectance of the green, red, and close and medium infrared best described the behavior of the biomass and carbon, due to the strong absorption of energy caused by the presence of vigorous, healthy vegetation, chlorophyll absorption, and vegetation moisture (^{Soria and Granados, 2005}). Table 3 shows the models selected with the stepwise regression procedure for estimating biomass and carbon based on spectral data.

Table 3 Selected regression models. 

B= Biomasa (Mg ha^-1); C= Carbono (Mg ha^-1); B2 = Reflectancia en banda verde; B3 = Reflectancia en banda rojo; NDVI = Índice de vegetación de diferencia normalizada.

Various studies have utilized spectral data to estimate the forest variables that have shown a wide variety of precision levels (RMSE) and determination coefficients (R²). ^{Aguirre et al. (2009)} calculated a carbon RMSE of 22.24 Mg ha^-1 in a Pinus patula Schiede ex Schltdl. & Cham. forest in central Mexico, using data from the SPOT 5 HRG sensor; they made estimates using multiple linear regression and the non-parametric k-nearest neighbor procedure (k-nn). Furthermore, 0.70 was the value of R² for the regression model selected to quantify carbon with a confidence level of 99 %.

On their part, ^{Labrecque et al. (2006)} determined RMSE values of 59, 59 and 58 (Mg ha^-1) for conifer, deciduous and mixed forests, respectively, in order to map the forest biomass in a pilot region in western Canada using Landsat TM imagery. ^{Escandón et al. (1999)} utilized the same sensor, but with a different approach, and evaluated two methods to estimate the arboreal biomass. The former method consisted in a six- band supervised multispectral classification; it used vegetation types identified by the makeup of the biomass of the dominant genera and by the estimated average height of the canopy. The latter method used normalized difference vegetation indices (NDVI) of bands TM4/TM3; TM4/TM5 and TM4/TM7; subsequently, they applied a regression model that related the average biomass with digital values or levels (DL) to the NDVIs. The exponential model turned out to have the best fit for the three NDVIs; the values obtained for the NDVIs were TM4/ TM3 (R²= 0.61), TM4/TM5 (R2= 0.67) and TM4/TM7 (R²= 0.68). Comparatively, the values obtained by ^{Aguirre et al. (2009)} are closer to those of the present study and are considered to be acceptable for this type of research.

The identified differences with respect to the previously cited experiences may be accounted for by the different spatial (^{Aguirre et al., 2009}) and spectral resolutions (^{Soria and Granados, 2005}) provided by the LANDSAT and SPOT satellites.

Comparison between the biomass-carbon estimates with the selected models

Figure 2 shows the relationship between the regression models developed as viable for use in forests. The utilized variable was the hydric stress index (NDVI43) because it proved to be highly significant (p≤0.0001) and had the highest correlation with the variables of interest. An inversely proportional relation was identified, i.e. the values for biomass and carbon were observed to increase with lower hydric stress indices (NVDI43), reflecting the carbon capture potential of this type of vegetation.

Figure 2 Observed and predicted biomass (Mg ha^-1) and carbon (Mg ha^-1) values, built with the NDVI43. 

Application and validation of the models

The selected models were applied to the reflectances of the bands and to the corresponding vegetation indices, and the biomass and aboveground arboreal carbon content were estimated by extrapolating them to the study area.

Table 4 shows the comparison between the values estimated for in-field biomass and carbon and those derived from the generated regression models.

Table 4 Biomass (Mg ha^-1) and carbon contents (Mg ha^-1), estimated in field vs. with the generated models. 

The availability of data from remote sensors with an enhanced spectral resolution can improve these assessments (^{Thenkabail et al., 2004}). Within this context, the obtained results coincide with the findings of ^{Fernández et al. (2005)}, according to whom, despite certain limitations, the images from this source are valid for estimating forest parameters at a regional scale.

Generation of biomass and aboveground arboreal carbon maps

Based on the analysis of the maximum biomass and carbon values of the sampling sites, numbers were determined for the pixels of the 400 Mg ha^-1 images of the biomass and for those of the 200 Mg ha^-1 images of the carbon content. In order to facilitate the interpretation of both maps, the data were grouped into classes and intervals (Table 5).

Table 5 Classes and ranges of values used to generate maps. 

Figures 3 and 4 show the generated maps of the biomass in the ejidos El Brillante and La Campana and the total carbon content of the plots in the ejidos La Ciudad and La Victoria.

In the maps generated for the forest regions of the study area, the various colors represent the different values of the biomass and aboveground arboreal carbon contents present in the assessed ejidos, in tons per hectare. For example, when taking the carbon map for ejido La Victoria as a parameter, each color refers to the carbon contents in this plot, with a maximum value of up to 200 tons in the forest areas.

Furthermore, this type of maps can prove useful to those who work with the natural resources, as a contribution to raising awareness of the causes of the loss of forests and forest carbon through the enhancement of forest management and the preservation of the forest’s biodiversity.

Figure 3 Total biomass maps (Mg ha^-1 ). 

Figure 4 Total aboveground arboreal carbon maps (Mg ha^-1 ).