To estimate the potential increase in biodiversity decline and the national level of conservation investment needed to counteract it in post-conflict Colombia, we used a model developed by Waldron et al.19. This quantitative model predicts national biodiversity status change, the biodiversity decline score (BDS), based on investment in conservation actions in relation to human development pressures. The model uses seven predictors related to the economy of each country, its biodiversity status or dynamics, and its conservation spending19.
Scenarios
We used the Waldron et al.19 model to predict (1) the expected increase in biodiversity decline immediately after the peace agreement (the post-conflict period), (2) the conservation funding needed to prevent this additional decline and (3) the investment necessary to avoid biodiversity decline. We used four scenarios to examine our questions.
The baseline scenario was the War BDS scenario, which estimated the BDS of the last 12 years of the conflict, before the peace agreement in 2016. Predictor variables related to human pressures were from 4–5 years before to appropriately represent the lag in the modelled effect19. We used the most recent available value of ‘strict-sense’ conservation investment19. The following three scenarios examined post-conflict options and were compared with this War BDS scenario.
The Peace BDS scenario predicted the BDS for a 12-year period post-conflict. The predictor variables related to human pressures were from the 11-year period immediately after the peace agreement. We assumed the same conservation spending as for the War BDS. The Lower BDS scenario estimated the necessary investment to achieve the War BDS. This represented a situation where the biodiversity loss during the conflict did not change post-conflict. For this scenario, we held the human pressure variables the same as in the Peace BDS scenario. The Prevented BDS scenario was exactly the same as the Lower BDS scenario, but we set a target of no biodiversity decline (BDS = 0).
We used the War and Peace BDS estimates to calculate the expected additional biodiversity decline post-conflict. Then, we used the model with data from the Lower BDS scenario to calculate the investment needed to prevent any additional biodiversity decline post-conflict. Finally, we used data from the Prevented BDS scenario to estimate the conservation investment necessary to halt biodiversity decline in the post-conflict period.
Data for predictor variables
We modified the predictors related to agriculture and economic growth to examine anticipated changes in human pressures. This revision allowed us to consider the expected agricultural expansion, in the form of percentage of agricultural land and growth, and economic growth, as the gross domestic product (GDP) and GDP growth. We also modified the function so that we could use it to estimate funding needs given a target BDS.
For the War BDS scenario, data on GDP, GDP growth, agricultural land area and agricultural land area growth were either available or easily computed. The data for GDP and the percentage of agricultural land from 2001–2012 were obtained from The World Bank28. The agricultural land growth was calculated as the difference between the percentage of agricultural land of consecutive years, and GDP growth was calculated from the GDP per capita data from The World Bank28.
For the Peace, Lower and Prevented BDS scenarios, we made projections about the predictors. For the GDP we used projections for 2017–2019, and for the GDP growth projections for 2019–2022 (ref. 33), and then selected an annual increase in the GDP growth of 0.3 percentage points for the remaining 5 years, corresponding to the most conservative estimate found in ref. 34. We then used our estimates of GDP growth for the whole time period to calculate the GDP per capita for the last 10 years, and used population projection to compute the GDP for the next 10 years.
To estimate the agricultural land and growth for the Peace, Lower and Prevented BDS scenarios, we used projections on deforestation. We developed our model to reflect the immediate consequences in agricultural expansion and deforestation post-conflict. Thus, we estimated the percentage agricultural land area using projected values of deforestation35. We support this approach based on two observations. First, at least 90% of deforested land was transformed to agriculture during past years36. Second, forest transformation to agriculture has been more aggressive since the peace agreement7,10,11. Thus, the processes that fuel agricultural conversion are stronger. For each year we added the deforested area to the previous agricultural land area. We then calculated the yearly percentage agricultural land area and computed the agricultural growth as the percentage difference between the agricultural land area of consecutive years. We took the minimum and maximum values of deforestation projections to create best- and worst-case scenarios.
We acknowledge that our use of the Waldron et al.19 model has limitations because we did not update all the predictors. Specifically, two ‘inertia’ terms that account for the effect of biodiversity decline occurring immediately before the time period of interest19. The coefficients associated with these terms have a positive effect on the BDS, which means that a more intense decline in the past will increase the predicted biodiversity decline. Given the increase in human pressures, the actual inertia terms are probably larger than the ones we used. Thus, the Peace BDS and the actual increase in biodiversity decline post-FARC may be larger.
The Model
To create a broad proxy for the expected cost of potential conservation interventions across Colombia, we estimated the OCC for agriculture at the 1 km2 scale. We estimated the OCC by building a spatially explicit probability model of forest conversion to agriculture and then paired it with the net present value of the expected return of different agricultural activities.
We calculated the OCC following the methodology proposed by Naidoo and Adamowicz24. Their approach models the expected net present value of potential net rents resulting from agricultural uses of a forested parcel, while accounting for the probability of conversion to agriculture. Provided that each agricultural use k has its own annual expected return per area of land Rk, and that each parcel i has a probability of conversion Pik from forest to agricultural use k, the expected value for a given discount rate δ is
$${{{mathrm{OCC}}}} = mathop {sum}limits_{i = 1}^{{I}} {mathop {sum}limits_{k = 1}^{{K}} {{{P}}_{i,k}} } frac{{{{R}}_k}}{{delta }}$$
(1)
Thus, the OCC of an area composed of several parcels is equal to the sum of the expected returns of the probable agricultural uses, weighted according to their probability of conversion, in each of the parcels, summed across all of the parcels.
We calculated the OCC for forested areas in three steps. First, we built a probability model to obtain the general risk of forest conversion (Pdef). Next, we built a second model that, given that a parcel had been transformed, predicted the probability of forest conversion to different types of agricultural activities (({{P}}_{{{{mathrm{ag}}}}_k})). We used both models to compute the total probability of conversion to each type of agricultural activity k in a parcel i (({{P}}_{ik} = {{P}}_{{{{mathrm{def}}}}_i} times {{P}}_{{{{mathrm{ag}}}}_{i,k}})). We then estimated the net present value of the expected return of each agricultural activity (Rk/δ) using literature and commercial prices and the costs of agricultural products.
Types of agricultural land use modelled
Our OCC model needed to represent relevant agricultural activities. Below, we justify our selection of three types of agricultural land uses: cattle ranching, coca crops and other crops.
Cattle ranching is expected to be a major driver of post-conflict deforestation11. This activity has accounted for 50% of deforestation, in the form of forest conversion to pasture, in past years36, and has considerably expanded post-conflict7.
Illegal coca crops are expected to be, and have been observed to be, an important driver of post-conflict deforestation12. This activity is at risk of increase where the withdrawal of FARC and the absence of state presence left a ‘power vacuum’ that facilitated other illegal groups gaining control of such crops in the territory7,11,12. Indeed, evidence shows that deforestation associated with coca cultivation increased as the conflict became less intense37.
Other crops were grouped into a single category with cattle ranching due to their small percentage contribution to forest conversion in our time frame (3%) compared with cattle ranching and coca crops (47 and 50%, respectively). We proxy for the extent of all other crops by using data on the distribution of three relevant agricultural products in the post-conflict period: cacao, oil palm and coffee. The cacao crop has high potential in most of the key post-conflict areas in Colombia, so it could have a major role in the peace transition38. Oil palm is important owing to its steep increase in cultivation during the last few years12, to the point that Colombia is now the largest producer in South America39. The relevance of coffee resides in its impact on the rural population, given that coffee crops are the only source of income for approximately 563,000 families and generates over 726,000 rural jobs40.
Landscape features data
We selected ten factors relevant to deforestation in Colombia to model the probability of forest conversion: proximity to roads, presence of FARC (binary: presence or no presence), population density, slope23, elevation, proximity to deforested areas, to rivers, to mining areas and to oil wells, and belonging to national and regional PAs10. National PAs restrict economic activities and are managed by the System of National Natural Parks, while regional PAs allow multiple-use activities and are managed by regional environmental authorities8,41. We did not include indigenous reserves or Afro-Colombian lands.
We used deforested areas from 1990 to 2000 from the Instituto de Hidrología, Meteorología y Estudios Ambientales (IDEAM)42, the water bodies map from the Department of Environment and Sustainable Development43 and maps from the Instituto Geográfico Agustín Codazzi (IGAC)44 to calculate the distance to already deforested areas, rivers, roads, mining areas and oil wells. The elevation map was obtained from NASA’s (National Aeronautics and Space Administration’s) Land Topography digital images45, and we calculated the slope using the elevation map. We computed population density as the mean value of the 32 mainland administrative departments from 2000 to 2012 using data from the Departamento Administrativo Nacional de Estadística46 (DANE; see Supplementary Table 3 for dataset details). We obtained a map showing the presence of FARC from the Fundación Paz y Reconciliación (PARES)47. All spatial data calculations were performed using software QGIS (https://www.qgis.org/en/site/, version 3.12.2) and R (https://www.r-project.org/, version 3.6.2).
Forest conversion and agricultural use model
We used a two-stage modelling process. First, we modelled the probability of an area being deforested by any driver (not exclusively due to agricultural expansion), using the total deforested area in the country in a 12-year period to parametrize our model (forest conversion model). Second, we modelled the probability that the deforestation was due to a particular agricultural activity (agricultural use model). To parametrize this second model, we used patches of land that were indeed transformed to an agricultural use in this same 12-year period. We combined these two models to obtain the probability that a patch of land was deforested to a particular agricultural activity.
We used a binomial logistic regression model to build our forest conversion model, which estimates the probability of forest conversion (Pdef). We used the land cover change from 2000 to 2012 across the country, available from IDEAM42, and reclassified each pixel cell as forested or transformed. We used the bayesglm function from the R arm package48.
For our agricultural use model, we built a second binomial logistic regression model to estimate ({{P}}_{{{{mathrm{ag}}}}_k}), the probability of conversion to each type of agricultural activity (cattle and other crops or coca crops) for a parcel that had been transformed. We employed data on forested areas in 2000 that had been converted by 2012. The coca crops cover map was obtained from the Sistema Integrado de Control de Cultivos Ilícitos (BIESIMCI)49. For the cattle ranching map, we used forested areas converted to pasture. Our other crop data contained temporary and permanent crops obtained from a land cover map43.
It should be noted that in logistic regression models, the probability of conversion does not change in a linear fashion, but the ratio of probabilities (odds) does. For the agricultural model, the odds describe the probability of conversion to coca crops over the joint probability of conversion to cattle and other crops. This implies that the variation between the probabilities, not the probability itself, changes constantly.
To check for spatial autocorrelation, we plotted spatial correlograms of the models’ residuals with Moran’s I. Because spatial patterns were present, we subsampled for pixel cells at a minimum distance of 20 km between points, which reduced the spatial effects adequately for our purposes, although it was most effective for the forest conversion model (Extended Data Fig. 1). We checked for collinearity in the predictor variables using variance inflation factor scores and removed the variables with a value >3 (distance to mines and oil wells; Supplementary Tables 4 and 5). We performed tenfold cross-validation to test the prediction accuracy of the models. This process splits the data into ten subsets and repeatedly fits the model with the data of nine of the subsets to compare its predictions with the remaining subset. We calculated the percentage of correct predictions (overall accuracy) each time and computed the mean as the final forecasting accuracy indicator.
Estimation of annual net rent
We estimated the net present values of the expected return of each agricultural activity to estimate the OCC of forested areas in Colombia. For cattle, we used annual net rent from a beef company50. The total annual net rent for other crops was calculated as the weighted average of the net rents for oil palm, cacao and coffee proportional to their land area in 2016 and 2017 (refs. 51,52,53). For coca crops, we used the average net profit for farmers who sell coca leaves54. We selected three discount rate values: 5, 10 and 20% (Supplementary Tables 6 and 7).
Predicting forest conversion and OCC
To predict the probability of forest conversion, we updated our spatial information on roads, deforested areas from 2007 to 2017 (ref. 42), FARC presence as the presence of FARC dissidents and deserters in 2017 (ref. 47), and population density as the mean population density by department from 2017 to 2023 (ref. 55). Together with the annual net rent for each agricultural activity, we used the probabilities of conversion of the two models to compute the OCC, or expected land value, of each forested pixel cell for the three discount rates using Eq. (1).
We recognize that the simplified national context of social violence when predicting the probability of forest conversion can limit the application of our results. Our models included FARC presence, and we used the presence of dissidents and deserters in this forecasting stage. However, this ignores other criminal groups that might influence the risk of forest conversion, particularly to coca crops, due to the ‘power vacuum’ left by the withdrawal of FARC and lack of state presence11. Because we overlooked the potential impact of other criminal groups, the probability of forest conversion, particularly to coca crops, could have been underestimated. This would imply an underestimation of the OCC in the areas with presence of these other criminal groups.
We used the rural cadastral values56 to validate our OCC results by comparing our predicted mean land values by administrative department in the country. Although rural cadastral values might not reflect the value of illegal coca crops, they were, to the best of our knowledge, the best available data for our purposes.
The STAR metric
The STAR metric is a measurement of the potential benefit to threatened and near-threatened species of actions aimed at reducing threats and restoring habitat20. The metric can be disaggregated spatially using the area of habitat for each species, showing the proportional potential contributions of conservation actions in particular regions. We focused on the STAR threat-abatement score (START) only. The START score can be further disaggregated by threat according to the contribution of each threat to the species’ risk of extinction, which allows analysis of potential abatement of species extinction risk by particular activities at particular locations. We took advantage of this trait and used the START metric in a specialized way, focusing on the threats posed by agriculture only on all the species with an area of habitat in Colombia. This resulted in 475 species considered (246 amphibians, 172 birds and 57 mammals), of which 169 are vulnerable, 124 near-threatened, 130 endangered and 52 critically endangered. Agriculture accounted for 52% of the total START. This focus on agriculture includes annual and perennial non-timber crops, wood and pulp plantations, and livestock farming and ranching, so we treated land converted to cattle and crops in the same way even though each land-use type has different impacts on species.
The use of the STAR metric has some limitations associated with the spatial distribution of the threat due to agriculture. First, the STAR metric is based on documented ongoing and expected future threats to the species according to the International Union for Conservation of Nature Red List. The majority of documented threats are ongoing, thus the majority of species threatened by agriculture are already being negatively impacted. This causes uncertainty in the assumption that avoiding further agricultural conversion will reduce species extinction risk, as additional activities to mitigate the impact of current agricultural activities on the species may also be required. Nevertheless, species assessed as threatened by agriculture are known to be vulnerable to this pressure, meaning that they would almost certainly suffer negative impacts under future agricultural expansion.
Second, there is uncertainty in the potential spatial distribution of agricultural expansion. Therefore, the STAR metric as we used it helped us identify sites with urgent potential benefits of avoiding agriculture. This could under-represent territories of great biodiversity value that are not currently impacted by agriculture, like the Amazon region.
Prioritization maps
We wanted to achieve a coarse methodology that could help decision-makers direct national conservation funding to the territories with the most potential benefits of halting forest conversion to agriculture. To pair the STAR scores with our modelled OCC, we divided the total range of STAR scores and OCC into high, medium and low values. Given the distribution of STAR scores, we divided the total range in the logarithmic scale. We classified each forested pixel cell into one of nine combinations of STAR scores and OCC. This analysis was later translated to the municipality resolution by calculating the mean STAR score and mean OCC of all forested pixel cells in each municipality, and applying the same classification system used at the pixel resolution. The distributions of aggregated STAR scores and OCC at the municipality resolution follow a similar pattern to the distribution by pixel cell, with small differences due to the grouping of the values in means (Extended Data Fig. 2b,c).
Reporting Summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Source: Ecology - nature.com
