Spatial Structure of Above-Ground Biomass Limits Accuracy of Carbon Mapping in Rainforest but Large Scale Forest Inventories Can Help to Overcome

Abstract

Precise mapping of above-ground biomass (AGB) is a major challenge for the success of REDD+ processes in tropical rainforest. The usual mapping methods are based on two hypotheses: a large and long-ranged spatial autocorrelation and a strong environment influence at the regional scale. However, there are no studies of the spatial structure of AGB at the landscapes scale to support these assumptions. We studied spatial variation in AGB at various scales using two large forest inventories conducted in French Guiana. The dataset comprised 2507 plots (0.4 to 0.5 ha) of undisturbed rainforest distributed over the whole region. After checking the uncertainties of estimates obtained from these data, we used half of the dataset to develop explicit predictive models including spatial and environmental effects and tested the accuracy of the resulting maps according to their resolution using the rest of the data. Forest inventories provided accurate AGB estimates at the plot scale, for a mean of 325 Mg.ha-1. They revealed high local variability combined with a weak autocorrelation up to distances of no more than10 km. Environmental variables accounted for a minor part of spatial variation. Accuracy of the best model including spatial effects was 90 Mg.ha-1 at plot scale but coarse graining up to 2-km resolution allowed mapping AGB with accuracy lower than 50 Mg.ha-1. Whatever the resolution, no agreement was found with available pan-tropical reference maps at all resolutions. We concluded that the combined weak autocorrelation and weak environmental effect limit AGB maps accuracy in rainforest, and that a trade-off has to be found between spatial resolution and effective accuracy until adequate "wall-to-wall" remote sensing signals provide reliable AGB predictions. Waiting for this, using large forest inventories with low sampling rate (<0.5%) may be an efficient way to increase the global coverage of AGB maps with acceptable accuracy at kilometric resolution.

Fig 1. Spatial distribution of inventory blocks from CTFT (1974–1976) and ONF (2006–2013).

Inventory blocks from CTFT (1974–1976) in pale grey polygons. Complementary inventory campaigns from ONF (2006–2013) in white circles (size represent the effective area covered by transects). Areas disturbed by harvesting or mining between 1974 and 2007 (in black) were removed from the dataset.

Fig 2. Flowchart followed for statistical analyses.

The grey colours indicate the different steps of analysis. Input data are represented in rectangles, analysis in ellipse and outputs in rounded boxes.

Fig 3. Variogram of biomass estimates from 500 m to 200 km according to distance classes.

The grey shape shows the confidence interval expected for each distance class under the null hypothesis (1,000 randomizations). The red squares indicate significant auto-correlation and the black circles imply no significant correlation.

Fig 4. Estimation of coefficient of variation and confidence interval of local mean according to grid resolution.

Inset boxplot shows CV of local biomass estimates according to grid resolution. Main part shows the fitted values confidence interval of the local mean according to the model log (IC_95%) = -0.59277–0.59128 * log (N) + 0.04988 * log (R) from N = 3 to 20 and R = 0.5 km to 8 km.

Fig 5. Coefficients of the selected GLM that predict biomass from environmental variables.

Grey bars and brackets indicate groups of modalities related to the different categorical variables. For continuous topographical and hydrographical variables (HAND, LOG, SLOpe and ALTitude) the coefficient value is multiplied by the mean of the variable.

Fig 6. Variogram of GLM residuals from 500 m to 200 km according to distance classes.

The grey shape indicates values expected under the null hypothesis of absence of spatial structure (1,000 randomizations). The red squares indicate significant auto-correlation and the black circles no significant structure. The dashed line represents the fitted exponential model used to predict the spatial error term.

Fig 7. Map of AGB (Mg.ha^-1) in French Guiana based on the complete model (KR).

Local darker or paler areas correspond to spatial error terms that can be modelled only within a short distance of the calibration plots and are null over most of the area.

Fig 8. Comparison of AGB values predicted by the different maps at 2-km resolution with test dataset.

Aboveground biomass (AGB) means at cell level for the test set are compared to the values predicted by the different maps at 2-km resolution: from the top left to the bottom right—KR, GLM, Baccini [15] and Saatchi [16]. The red line indicates the 1:1 relationship (expected slope). The size of the circles indicates the number of plots for each cell in the test set (from 3 for the smallest to 12 for the biggest).