Agriculture-driven deforestation in the tropics from 1990–2015: emissions, trends and uncertainties

The terms error and uncertainty, are often used interchangeably (Taylor 1997, Heuvelink 2005), and in this study both are used; error to represent the difference between an estimate and the true value, and uncertainty to represent a quantification of the distribution of error. If uncertainty information is not available, potential error sources contributing to uncertainty were identified and the magnitude of each source estimated for each input dataset. All datasets were assessed for the broad causes of uncertainty described by the IPCC (2006) (summarized in table 3).

For each country, the uncertainties identified in each dataset are defined for each time period (table 4).

For many datasets,it was not possible to quantify uncertainty using data (e.g. with statistical approaches), so data from additional sources and expert judgement were used (appendix S2).

The overall uncertainty (σ²) was calculated by combining the uncertainties associated with each error source (σ_i²) (i...n) (equation (6)). Where all terms (error sources) are assumed to additively contribute to total uncertainty

$$\begin{equation} \sigma ^2 = \sigma _i^2 + \sigma _{ii}^2 + \cdots + \sigma _n^2. \end{equation} \tag{ 6 }$$

Following guidelines from the IPCC, upper and lower estimates correspond to a 95% confidence interval (CI), as a percent of the mean (IPCC 2006) (equation (7))

$$\begin{equation} {\rm CI} \pm = \sqrt {\sigma ^2 } \cdot 1.96. \end{equation} \tag{ 7 }$$

In case the estimate for the area of deforestation (A) was zero for a particular country, when calculating the confidence interval, A was substituted by the mean deforestation estimate from other (non-zero) time-periods in that dataset.

Once errors have been calculated for each dataset, they must be propagated to derive uncertainties for D and ADD. Errors associated with the inputs were assumed to be independent, normally distributed, and without bias. We used the exact equation for the variance of the product of three and four independent random variables (Goodman 1962) (equations (8) and (9)) to calculate the output variance (σ²)

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\sigma ^2 (D)} \hfill & = \hfill & {\sigma ^2 [A \cdot {\rm CB} \cdot f_{\rm CBLU} ]} \hfill \\ {} \hfill & = \hfill & {(\sigma ^2 (A) + A^2 ) \cdot (\sigma ^2 ({\rm CB}) + {\rm CB}^2 )} \cdot \hfill \\ & {} \hfill & (\sigma ^2 (f_{\rm CBLU} ) + f_{\rm CBLU} ^2 ) { - A^2 \cdot {\rm CB}^2 \cdot f_{\rm CBLU}^2 } \hfill \\ \\ \end{array} \end{equation} \tag{ 8 }$$

$$\begin{equation} \begin{array}{l@{}l@{}l@{}} {\sigma ^2 ({\rm ADD})} \hfill & = \hfill & {\sigma ^2 [{\rm A} \cdot {\rm EF} \cdot f_{\rm CBAgri} \cdot f_{\rm AAgri} ]} \hfill \\ {} \hfill & = \hfill & {(\sigma ^2 ({\rm A}) + A^2 ) \cdot (\sigma ^2 ({\rm CB}) + {\rm CB}^2 ) \cdot (\sigma ^2 (f_{\rm CBAgri} )} \hfill \\ {} \hfill & {} \hfill & { + f_{\rm CBAgri}^2 ) \cdot (\sigma ^2 (f_{\rm AAgri} ) + f_{\rm AAgri}^2 )} \hfill \\ {} \hfill & {} \hfill & { - {\rm A}^2 \cdot {\rm CB}^2 \cdot f_{\rm CBAgri}^2 \cdot f_{\rm AAgri}^2 }. \hfill \\ \end{array} \end{equation} \tag{ 9 }$$

The contribution of each component was assessed as the reduction in the output variance when the corresponding input variance was set to zero. In other words, we recalculated (equations (8) and (9)) multiple times, setting for each recalculation one of σ² (A), σ² (CB), σ² (f_CBLU) or σ²(f_CBAgri), σ²(f_AAgri) to zero. The lowest outputvariance for which the variance of that element (A, CB, f_CBLU or f_CBAgri, f_AAgri) is set to zero, is the one with the largest contribution to the uncertainty.

Some of the uncertainty estimates were based on expert judgment. We used a sensitivity analysis to explore how a different judgement would alter the uncertainty of the final value. Initial assumptions (section 2.1), were compared to three adjustments (table 5). For each adjustment, the emissions are recalculated, and the change in emissions estimate uncertainty calculated.

We usedall available datasets to calculate emissions giving higher weight to the most certain dataset available. Figure 2 shows the weight given to the datasets for each component, based on their uncertainty.

Zoom In Zoom Out Reset image size

Figure 2 is driven by both the availability of data, and the weight of each dataset. For component A, RSS and FRA have the highest weighting for the most countries (33 each). Large countries (Brazil, India, China) tend to have a higher weighting for the RSS data, since they have a larger number of sample units, which increases the certainty of the data. The weighting of the RSS data is also driven by the variability in the proportion of deforestation in each sample within a country. Countries which have better data (indicated by higher capacities, or access to better data and support), for example the majority of countries in South and Central America, have a higher weighting for the FRA data. In Africa, more countries (20 out of 43 countries) have the highest weighting for the Hansen data. These tend to be countries with lower capacity, and relatively few RSS sample units. Kim consistently gets low weightings, as although the uncertainties in terms of the CI when expressed as a percent of the estimation are the same as Hansen, the estimates tend to be higher, so variances are also higher. For the biomass in forests, these datasets generally have a more equal weighting, however the Avitabile dataset is more certain, and carries the highest weight in 85 of the 91 countries. For the remaining 6 countries, Saatchi has the highest weight. For the fraction of deforestation which is driven by agriculture, in 43 out of 91 countries only the Hosonuma data are available. For the remaining countries, most (42 of 48) have a higher weight for the Hosonuma data. The countries where De Sy2 has a higher weight are larger countries, which as a result have more sample units (Brazil, India, Mexico, Namibia, Venezuela and Zambia).

The area of deforestation is dynamic and drives most of the trend of the emissions calculations. Between the first two time periods (1990–2000 and 2000–2005), RSS more frequently has the highest weighting. The last two trends (between the last three periods) show a more even distribution, where FRA and Hansen are almost equally likely to contribute the most to the estimate of deforestation. However FRA tends to dominate in Latin America and Asia, while Hansen dominates in Africa, with the exception of some countries including Democratic Republic of Congo, South Africa, and Madagascar most notably (figure 3, table 6).

Zoom In Zoom Out Reset image size

The trends for the same time period derived from two different datasets (either FRA and RSS or FRA and Hansen) can disagree in the direction or the magnitude of the change. This difference can be relatively large (mean absolute trend difference is 1.47), with the FRA and RSS being on average more different (absolute difference of 1.9 between the periods 1990–2000 and 2000–2005), and FRA and Hansen being more similar. In many cases, the trend direction is the same for both datasets, and the trend between the last time periods (2005–2010 and 2010–2015) has the most agreement from the contributing datasets, with only 16% of countries having conflicting trends. For the trend between the first two time periods (1990–2000 and 2000–2005) 28% of countries have conflicting trends from FRA and RSS.

A number of countries in Asia have the highest agriculture-driven deforestation between 1990 and 2005. In Africa, the majority of countries had their highest period of agriculture-driven deforestation in 2010–2015 (figure 4).

Zoom In Zoom Out Reset image size

Latin America is the greatest contributor to global emissions from deforestation and agriculture-driven deforestation, while Africa is the lowest but shows a continual growth in emissions (figure 5). In Africa, the highest emission rates (412 ± 75 Mt CO₂ yr⁻¹) occur in the 2010–2015 period, whereas for Latin America the emissions peaked during 2000–2005 (971 ± 148 Mt CO₂ yr⁻¹) and decreased afterwards. Overall, the highest annual emission rates (1792 ± 133 Mt CO₂ yr⁻¹) occurred in 2005–2010, with the lowest in 1990–2000 (1511 ± 174 Mt CO₂ yr⁻¹). Although Brazil's emissions have declined since 2005, these emissions still dominate in the region. Emissions from agriculture-driven deforestation are on average 72% of emissions from all deforestation. This is highest in Latin America (78%), and lowest in Africa (62%). In Asia it is 67%. The remainder of Latin America has the highest proportion of agriculture-driven deforestation (94%), and Indonesia has the lowest (52%).

Zoom In Zoom Out Reset image size

The uncertainty of the emissions estimates of agriculture-driven deforestation range between ± 24.9 to ± 283.1% of the estimate, with a mean of ± 62.4% per country (figure 6). The uncertainty for estimates of emissions from deforestation range from ± 10.7 to ± 260.9%, with a mean of ± 29%. At the country level, typically Latin America has lower uncertainties than Africa or Asia, however the highly forested countries in Asia and Africa (for example Indonesia and Democratic Republic of Congo) also have lower uncertainties. Uncertainties for agriculture-driven deforestation emissions are higher than uncertainties for deforestation emissions (figure 5). When country uncertainties are aggregated, uncertainties for emissions are higher in Asia ( ± 22.5% for ADD and ± 7.7% for D), followed by Africa ( ± 16.7% for ADD and ± 6.1% for D), and Latin America ( ± 15.9% for ADD and ± 6.1% for D). Uncertainties for global aggregates are ± 11.4% for ADD and ± 4.3% for D.

Zoom In Zoom Out Reset image size

For estimates of emissions from deforestation, A is more frequently the largest contributor to the uncertainty, and f_CBLU is more frequently the smallest contributor. In the case of emissions from agriculture-driven deforestation, f_AAgri is more frequently the largest contributor to the uncertainty, and CB is more frequently the smallest contributor. This pattern is clear in all continents, with some exceptions, including a number of countries in Latin America, wherethe forest biomass fraction and lost biomass fraction contribute to a relatively large amount of the uncertainty (figure 7).

Zoom In Zoom Out Reset image size

Sensitivity analysis was implemented to determine how changes to the uncertainty estimates for each input variable influence the uncertainty of the emissions estimates. Increasing the estimated uncertainty associated with the adjustment for the thematic mismatch leads to the largest change in the uncertainty of the emissions estimate, and increasing the extrapolation error leads to the smallest change to the uncertainty. A 10 fold increase in the uncertainty related to the adjustment of the thematic mismatch or bias leads to a ± 17.74% change in the final uncertainty estimate, which accounts for 63.73% of the original uncertainty for emissions from agriculture-deforestation. Doubling the extrapolation error from ± 15% to ± 30% leads to the smallest change in uncertainty for deforestation emissions; only ± 0.37%, which accounts for 1.5% of the original uncertainty estimate (appendix S2).

In general, our results for area of forest loss are similar (73–108%) to other published data (see appendix S3 for details). All the studies used in the comparison use a land cover definition, and thus report forest losses in land where there is no change in use, which would not be included in our study. Our emissions from deforestation are 52–75% lower than those from other studies because we take into account the new biomass which replaces forest biomass, or the biomass from the forest which remains on the land, whereas other studies assume that all the biomass is lost following deforestation. This makes a large difference in Asia, where due to tree crops often replacing forests, the biomass lost is replaced by more new biomass than in other land use conversions. The definition of forests can also explain some of the differences. Tyukavina et al (2015) have a higher canopy cover threshold (25%), while Harris et al (2012b) look at the removal of any forest cover. This particular difference would lead to an underestimation from our study in comparison to Tyukavina et al (2015), and an overestimation in comparison to Harris et al (2012b). Achard et al (2014) distinguishes between forests and tree cover mosaics (>70% and 30–70% tree cover respectively), and also 'other wooded land' (OWL). OWL is defined as 'all other woody vegetation (height <5 m), including shrubs and forest regrowth'. OWL is also likely to be found in dry forests, which are difficult to measure, and there is more disagreement over their extent (Bastin et al 2017). Data from Achard et al (2014) used for the comparison could include OWL which is not included in the definition of forests in this study, so we would expect this to lead to higher results from Achard et al (2014). The results for South and Central America are more in agreement in these two studies than those for Africa and Asia. Proportionately, according to Achard et al (2014) there is more OWL in Africa: 58% of the total forest in 2010, in comparison to South and Central America and South East Asia, which have 18% and 31%, so this could explain the differences.

The best combination of data for emissions estimates differed from country to country (figures 2 and 3). Input data were weighted according to their uncertainty, and these weights can guide decision makers on data selection in similar contexts to this study. It should also be noted that in addition to the uncertainty of the dataset influencing the weight, the magnitude of the estimate also influences the weight. For example, the Kim dataset consistently had lower weights than Hansen, even though the two datasets had similar percent confidence intervals, due to the higher estimates of Kim. For biomass, the Avitabile dataset is most often selected, unless it is not available, where Saatchi is selected. However all datasets have similar uncertainties so could all be considered useful. For the agriculture-driver fraction, the results of this study suggest that only large countries should use De Sy2 above the Hosonuma data. We however suggest that an individual examination of both datasets at the national level may lead to different conclusions about the reliability of the datasets. In fact, both Hosonuma and De Sy2 were found to have large uncertainties.

The urgent need to limit global temperature increases below 2 °C will require actions to reduce all emissions sources, and as a major source of deforestation emissions, reducing agricultural expansion into forests should be considered as a mitigation priority. Latin America currently has the largest emissions from the regions in this study, however emissions have been reducing over the period of the study (figures 4 and 5). Africa has seen a consistent increase in emissions between 1990 and 2015, as predicted by past studies (Barnes 1990). Countries or regions with the largest increase in emissions could be targeted for mitigation actions, for example DRC which saw a large increase between 1990 and 2005, and humid Africa which saw a large increase between 2005 and 2015. In order to address agriculture-driven deforestation, targeted interventions should be developed which address specific agricultural activities. In these cases, interventions in the agriculture sector to mitigate emissions from agriculture-driven deforestation can be effective. Areas with either a yield gap which can be reduced, or with available degraded land which can potentially be rehabilitated have been highlighted for their mitigation potential (Carter et al 2015). There is some debate on the conditions under which these agricultural interventions will be successful (see for example Angelsen and Kaimowitz 2001), however implementation of forest protection activities (such as REDD+) are highlighted as being essential to ensure that forests are protected (Mertz and Mertens 2017). Although in all the countries included in this study, agriculture was found to be the largest driver of deforestation, it could be the case that in the future other drivers become more important, and we acknowledge the need to monitor all drivers. In this paper we do not address emissions related to other carbon pools (such as soil), or indirect emissions for example those related to the life cycle of agricultural products (which may lead to further deforestation). These will result in additional emissions above those considered in this study. Additionally we do not include emissions from forest degradation, which may be significant as estimated by (Baccini et al 2017), and can also result from agriculture.

Estimates of the trend in area of deforestation were provided by two datasets for each time period, and were in some cases very different (section 3.2). We used a weighted mean to prioritize the most certain estimate, however in the case that the two datasets predict an opposite trend (one increasing and one decreasing), the weighted average will thus produce a trend which is closer to 1, which may not reflect the actual trend. This effect will be most seen in the first time period, where there is more disagreement between datasets. Another challenge in this study was the production of comparable uncertainty estimates. Using uncertainties from the data providers themselves, could lead to better results. The current method relies on assumptions about the uncertainties related to the datasets, as many were lacking information or had uncertainty information which could not be used in this study. Our research estimated uncertainties which aimed to capture all the sources of error for each dataset, and it could be that errors exist which were not included in the study. In some cases, expert judgement was used to quantify the uncertainty, which may be erroneous. However sensitivity analysis confirmed (section 3.4) that in many cases the change in uncertainty was not large following a change in the assumptions which were based on expert judgement. Hence, if our uncertainty estimates based on expert judgement are incorrect, this will not substantially influence the overall uncertainty estimates. Additional uncertainties exist in the ancillary data are used to harmonize the datasets (due to thematic mismatch) but we chose not to include those uncertainties, as they are unknown, although some error is assumed to be included during harmonization. Future studies could explore this further. Lack of available data also limited the study. Only the area dataset was considered to be dynamic, with the remaining datasets assumed to be constant over time. Because there is not available data over time for emissions factors or drivers of deforestation, we were not able to capture this dynamic in our end results. This means that the trend in emissions is determined mainly by the area data, and in reality it may be influenced by changes over time in the other inputs, for example emissions factors.

Uncertainties associated with our estimates of D and ADD are in some cases very high at the country level (for example Uruguay is ± 182% for ADD), although the average for ADD is much lower (± 62%) (figure 6). Large uncertainties are in line with findings in Houghton et al (2012), and Roman-Cuesta et al (2016) who found that uncertainties from forest loss contribute to 98% of the uncertainty of AFOLU emissions, while only contributing to 69% of the emissions. The authors recommend that uncertainties are reported in future datasets (to increase transparency), and that improvements in datasets (increased certainty) should be pursued. Since area data and agriculture-driver factors are the least confident, improvement in the uncertainties related to these estimates will also provide the greatest reductions in uncertainties of emissions estimates. There are two ways to address these uncertainties when reporting for mechanisms such as REDD+. Either the upstream uncertainties are reduced by improving the input datasets, or the emissions estimates are adjusted downstream, by discounting or reducing the estimates (a conservative approach) to avoid overestimating emissions reductions (Pelletier et al 2015). The findings of this study suggest that the upstream adjustments should be made to avoid having to implement downstream adjustments, which reflect negatively on the credibility of the mechanism.