Selective logging is one of the leading causes of forest degradation in the Brazilian Amazon region. The Brazilian Federal government has adopted a forest concession policy as a strategy to mitigate impacts of selective logging and regulate operations of the tropical timber industry in Brazil. This study used fractional forest coverage derived from satellite imagery and field data to assess forest degradation in two selectively logged study sites within the Jamari National Forest, a protected area located in the western Brazilian state of Rondônia. Initially, we estimated the fractional coverage from vegetation indices using RapidEye imagery and compared to gap fraction data derived from hemispherical photos acquired in the field. Subsequently, we estimated the impacts of different types of selective logging activities (log decks, primary and secondary roads, tree fall gaps, and skid trails) on forest cover using the fractional coverage dataset. The NDVI showed the highest R2 (0.56), indicating that 56% of the sample variation in fractional coverage derived from ground measurements can be explained by fractional coverage derived from the NDVI model. Our results also showed that the intensity of canopy impacts may vary according to the selective logging activity, ranging from skid trails to log decks which had the lightest and the heaviest canopy impacts, respectively.
Several efforts to improve the sustainability of forest resources exploitation have been made throughout the world (
In 2006, the Brazilian federal government adopted a new national forest policy to respond to a myriad of social, economic, and environmental pressures placed on forests of the Brazilian Amazon (
Despite the ongoing conservation efforts, selective logging is one of the main drivers of the extensive forest degradation occurring in the Brazilian Amazon (
As a result, a more accurate and precise way of detecting selectively logged forests is needed to understand the spatial distribution of logging activities and to estimate their overall impacts (
Remote sensing is an important data source for the improvement of command-and-control tools, mostly because of its proven capacity to provide data relating to large areas of continuous forest (
Several authors have already addressed the existing technical challenges that hamper the efforts to detect and to map areas disturbed by selective logging activities.
Despite these technical difficulties, remote sensing has provided some estimates of the extent and intensity of selectively logged area in the Brazilian Amazon Basin (
This research aimed at improving the currently available analytical tools used to estimate the impacts of selective logging on tropical forests. To achieve that, we combined field data and RapidEye imagery to estimate the impact of selective logging at two study sites within the Jamari National Forest (JNF), located in the western Brazilian Amazon state of Rondônia. First, we estimated the fractional coverage (
The Jamari National Forest (JNF) was selected because it was the only federal forest concession operating at the time we collected our data. Our study was designed to complement the tests related to the establishment of a monitoring system of forest concessions, which have been conducted by the Brazilian Forest Service.
The JNF was created by Federal Decree 90224/1984 as a protected area for sustainable use only. The territory of the JNF includes parts of the municipalities of Itapuã do Oeste, Cujubim, and Candeias do Jamari, which are located in the Western Brazilian Amazon state of Rondônia. The JNF encompasses a total of 220 000 ha of tropical forest. The main vegetation type found at the JNF is the dense tropical forest, but there are also patches of open tropical forest (
In 2008, the Brazilian federal government allocated 96 000 hectares of the JNF for forest concession, which included three management units. Our study sites were the first Annual Forest Production Plots (AFPP) of Forest Management Units (FMU) 1 and 2 (
The total forest area under concession in the study sites encompasses 1662 ha (AFPP in FMU 1 encompasses a total of 594 ha, and the AFPP in FMU 2 a total of 1068 ha). According to production reports from the Brazilian Forest Service, it was harvested in 2010 and 2011 an average of 14.8 m3 ha-1 and 9.6 m3 ha-1 within the study sites (FMU 1 and FMU 2, respectively).
Gap fraction data was acquired during a fieldwork conducted from October 03 to 15, 2011. Sampling points included disturbed forest by selective logging activities (tree fall, skid trails, log decks, and primary and secondary roads). Additionally, gap fraction data was acquired for unlogged forests and used for control data in the analysis. These sampling techniques were previously used in gap fraction data collection by
The sampling scheme used in the different types of selective logging disturbances included the following steps: (a) measurement of one third of log decks in both study areas; (b) measurement of patches of primary and secondary roads in each AFPP, by randomly selecting transects of 50 m length. The photos shots were taken at a 10 m interval along each transect; (c) measurement of skid trails using transects of 50 m length to avoid tree gaps and log decks; (d) forest canopy disturbances by tree fall gaps were measured using randomly selected trees (the same number as that of road patches), in the transects of 50 m length starting at the tree stumps. Two transects of 1000 m length were used to sample undisturbed forests, one at each study area. In this case, photo shots were taken at 60 m interval along each transect.
Hemispherical photographs were taken in the field using a 3.2 megapixel camera coupled with a fisheye lens and a vertical-horizontal leveler. These photographs were taken under favorable weather conditions and avoided the incidence of direct solar radiation. All sampling points were georeferenced using a handheld GPS unit (Garmin GPSMAP® 76 Cx) with locational accuracy of < 10 meters.
Gap fraction was estimated using the Gap Light Analyzer (GLA) software (
This study used RapidEye images, orthorectified at 3A level with 5 m spatial resolution. Images were acquired on September 21, 2011; this date was selected because it was the closest to the fieldwork period. A cloud mask was applied to images with substantial cloud cover. However, images of some parts of the study area were acquired in May for two reasons: first, because at this time there was a lack of cloud cover, and second because most of the marketable trees at those parts had already been harvested in the previous year.
RapidEye imagery was radiometrically corrected using the Radiometric Scale Factor available in image metadata, then converted to top of the atmosphere reflectance using earth-sun distance and exo-atmospheric irradiance values for each band (available from RapidEye Product Specification). The ERDAS® 9.1 software was used to perform this calibration. Values of solar distance and the zenith angles applied in the reflectance conversion are shown in
Three vegetation indices were tested to estimate the fractional forest coverage, and these are summarized as follows. The NDVI (
The Modified Soil Adjusted Vegetation Index (MSAVI) is a modified version of the Soil Adjusted Vegetation Index (SAVI) developed by
where
where
In addition,
where
where
A mixture model was used to derive the fractional forest cover based on the vegetation indices. In this model, the reflectance of each pixel was defined as the sum of the individual reflectance values of all components weighted by its proportion in the total coverage (
According to
and it can be rewritten as (eqn. 6):
where
The values of the vegetation indices for the two endmembers employed in the model were determined from histogram analysis and visual inspection of VI images. A pixel value from a deforested area and the average value of vegetation in undisturbed forests were used for bare soil and vegetation values, respectively. When a linear model with more than two components is used, eqn. 6 cannot be considered because the analytical model assumes that a pixel can only consist of soil and vegetation components.
A fractional coverage (
Each location point acquired in the field was associated with a specific type of selective logging environment and, the gap fraction data derived from hemispherical photos were then linked with the attribute table of each point location. One hundred and sixty four points for disturbed and undisturbed forests were used in the analysis. A 10-m buffer zone was generated around each point to estimate the arithmetic mean of pixels values from fractional coverage images within this area. These buffer zones were created to minimize the uncertainty related to the positional accuracy of the GPS field point locations and image geometric corrections.
Gap fraction measurements of 164 field sample points and their associated fractional coverage values derived from vegetation indices were used as input for the statistical analysis. As a result, the contribution of each type of forest disturbance by selective logging compared to undisturbed forests was assessed.
The relationship between the fractional coverage retrieved from vegetation indices and the gap fraction measured in the field using hemispherical photos were tested using a linear equation model. The best fit equation was employed to select the optimum vegetation index. Additionally, we applied the criteria suggested by
Canopy cover impacts by selective logging were assessed using multiple regression analysis. The fractional coverage estimated from the best vegetation index was used as the dependent variable, and the location of the five types of logging activities were used as the independent variables. Field measurements of undisturbed forests spatially located contiguously to disturbed forests were used as control. Multiple regression analysis is a useful technique for estimating the partial effects of independent variables because it controls other factors that could simultaneously affect the dependent variable (
where
The null hypothesis were defined as: each independent variable (log decks, primary roads, secondary roads, tree fall gaps, and skid trails) used in the regression model has no effect on fractional forest cover. A 95% confidence level was adopted. Simple and multiple linear regression analyses were performed using the R statistical package, version 3.0.1.
The simple linear regression analysis showed the best relationship between fractional coverage derived from NDVI and gap fraction measured using hemispherical photos.
NDVI’s least residual dispersion between the estimated and observed values, R² value, absolute and relative standard errors, and correlation coefficient between the observed and estimated values confirmed this index as the most accurate of the three tested. As a whole, the NDVI model outperformed to estimate forest fractional cover when compared to MSAVI and GEMI.
Complementarily, the validation analysis confirmed that NDVI was the best index as input of a linear mixing model to estimate canopy openness.
Based on these overall results, the NDVI was selected as the most accurate vegetation index to estimate canopy openness for our study sites.
The existing gradient among different type of selective logging environments was assessed by applying a multiple linear regression model using the fractional forest coverage estimated from the NDVI as the dependent variable. The statistical results obtained from our regression model were: R² = 0.58; standard error of the estimate in units of the estimated variable = 11.55; standard error of the estimate in percentage = 16.30; correlation between observed and estimated values = 0.77.
Forest cover was estimated as percentage values and our results are very explicit regarding the impact (estimated losses of canopy coverage) of selective logging on the forest. At a 95% confidence level, the contribution of log decks and primary roads in the decrease of forest canopy cover were significant, while the effect of secondary roads, tree fall gaps and skid trails were not (
The multiple regression model showed a coefficient of determination of 0.58, attributed to the high variation present in the data. The standard error of the estimate (percentage) was approximately 16%. The forest canopy environments that sustained the most damage (
The NDVI index showed the best performance to estimate forest canopy cover compared to the MSAVI and GEMI. Despite the initial expectation that correcting the atmosphere and soil effects would benefit the performances of both GEMI and MSAVI, our results did not confirm this hypothesis. The essence of the mapped targets may partially explain these results. Roads, log decks and tree fall gaps opened during logging activities reveal small portions of bare soil beneath the dense forest canopy. It is likely that, in our case study, the soil signal coming from these features improved the performance of NDVI.
Selective logging activities cause different degrees of canopy openness (
Further selective logging activities cause less canopy damage, such as tree fall gaps, secondary roads and skid trails, as most of their impact occurs at the ground level. As a result of those intrinsic characteristics and the adoption of reduced impact logging techniques in Brazilian forest concessions, these activities caused cryptic forest degradation that could not be properly detected by our remote sensing approach. Optical remote sensing can detect selective logging based on contextual elements (network of roads and log decks and associated canopy damage), but it has limited capacity to distinguish the structural changes that occur below the canopy (
The adjusted regression model used for selecting a vegetation index for deriving forest cover indicated that 56% of the variation in fractional canopy cover estimated with field data can be explained by variations in the fractional forest cover estimated from the NDVI. However, the coefficient of determination needs to be evaluated with great caution and whenever possible, along with other criteria, as recommended by
The relationship between fractional forest cover estimated with field data and that obtained with remotely sensed data is not linear. With higher canopy openness, the difference between the two data sources is quite large, mainly because of the wide angle of the fisheye lens that shows vegetation pixels at the image edges even in open areas. Satellite images do not have this constraint, especially those with higher spatial resolution, which results in a reduction in the spectral mixture in the pixels.
We minimized potential autocorrelation among samples by adapting our sampling scheme. We accepted, however, some degree of spatial autocorrelation (Global Moran’s
Although a buffer zone surrounding the georeferenced field points was used to reduce uncertainties, we acknowledge that the resolution of the remote sensing data is coarser than the GPS locational accuracy. The accuracy of field data has been previously reported as negatively influencing predictions of forest attributes (
The intensity of impacts caused by selective logging in tropical forests is usually related to the adopted harvesting techniques and intensities, with different levels of ground and canopy structural damage, as well as the subsequent losses of biomass and the forest recovery time (
In this study, we estimated the specific contribution of logging operations to damages occurring to the forest structure. Our results showed that log decks and primary roads were responsible for the greatest impacts on forest canopy. Based on this result, we emphasize that an appropriate forest planning is required before the beginning of selective logging operations to minimize the impacts caused by those forest activities. Additional care should be taken during the implementation of forest management plans (secondary roads construction, tree felling and skidding).
Recent studies (
Despite the fact that our findings are derived from a single case study in the Western Brazilian Amazon, we believe that our results can be extrapolated to other tropical ecosystems that are currently experiencing selective logging activities. This assertion is supported by the accuracy of the techniques we used to assess canopy cover variations in tropical forests under timber exploitation. Nonetheless, we recommend additional research to be conducted for the improvement of the methods used to quantify selective logging impacts on tropical forest canopy. Potential areas of investigation could include the use of additional sensors and different forest types, and the assessment of alternative selective logging strategies to be used in public forest concessions and private projects.
Finally, our findings may contribute to improve the monitoring tools and policies of forest concessions conducted by governmental agencies and civil society organizations in tropical regions.
The authors are grateful to the Brazilian Forest Service for providing access to the forest concession area and logistics support for the fieldwork, and to the National Institute for Space Research for financial support for the fieldwork and image acquisition.
ERP conceived the study, carried out field work and image processing, performed data analysis and wrote the paper; EATM conceived the study and wrote the paper; FAL performed data analysis; and MAP wrote the paper.
Study site location: the Jamari National Forest, its Forest Management Units (FMU), and Annual Forest Production Plots (AFPP).
Example of: (a) hemispherical photograph of a log deck; and (b) binary image resulting from semi-automated image processing using the Gap Light Analyzer.
Flow diagram of the relationship between gap fraction from field data and fractional coverage derived from vegetation indices.
Residual dispersion of adjusted (a) and validated (b) models for estimation of fractional canopy coverage, retrieved from NDVI, GEMI , and MSAVI.
Distribution of fractional coverage derived from NDVI (left) and residual distribution (right) per each selective logging activity. (LD): log decks, (PR): primary roads, (SR): secondary roads, (TG): tree fall gaps, (ST): skid trails, (UF): undisturbed forests.
Acquisition date and the geometry of RapidEye images acquisitions used in this study.
RapidEyeTile | Acquisitiondate | Solardistance (AU) | Zenithangle |
---|---|---|---|
2034916 | May 05-2011 | 1.01299 | 32.58 |
2034915 | September 21-2011 | 1.00402 | 11.37 |
2034814 | September 21-2011 | 1.00402 | 11.66 |
2034815 | September 21-2011 | 1.00402 | 11.54 |
Statistics of adjusted and validated Fractional Coverage models. (
Model | R² | Syx(adj) | Syx (%)(adj) | ryx(adj) | Syx(val) | Syx (%)(val) | ryx(val) | ||
---|---|---|---|---|---|---|---|---|---|
FC NDVI | 64.82 | 0.25 | 0.56 | 4.04 | 5.68 | 0.75 | 5.38 | 6.47 | 0.71 |
FC GEMI | 63.12 | 0.29 | 0.48 | 4.41 | 6.52 | 0.69 | 5.57 | 6.74 | 0.68 |
FC MSAVI | 72.25 | 0.21 | 0.51 | 4.31 | 8.37 | 0.71 | 5.55 | 6.72 | 0.68 |
Coefficients of multiple regression model for the fractional forest cover derived from the NDVI.
Parameter | Estimate | Std. error | p-value | |
---|---|---|---|---|
Intercept | 83.9 | 2.2 | 38.4 | 0.00 |
Log decks | -33.2 | 2.8 | -11.9 | 0.00 |
Primary roads | -20.6 | 3.4 | -6.1 | 0.00 |
Secondary roads | -3.2 | 2.9 | -1.1 | 0.27 |
Tree fall gaps | -6.2 | 3.5 | -1.8 | 0.07 |
Skid trails | -0.2 | 3.5 | -0.0 | 0.96 |