Scenarios of land use and land cover change in the Colombian Amazon to evaluate alternative post-conflict pathways

Study area

In Colombia, the Amazon region represents 42.3% of the territory with an estimated area of 483,164 km². In this area, 14% is dominated by agricultural lands, secondary vegetation and fragmented forests. Currently, 86% of the area corresponds to natural areas in a good state of conservation, where forests are the dominant coverage⁶. In the northwest area, the region borders the Andean Cordillera and Orinoquía to the north. The political-administrative division includes the departments Amazonas, Caquetá, Guainía, Guaviare, Putumayo and Vaupés, and part of the departments Cauca, Meta, Nariño and Vichada. The human population is estimated at ~ 1.4 million, with a density of 2.5 inhab/km². Internal conflict and poverty make this region one of the most important population dynamics in the country in terms of displacement³⁶. The geographical location of the study area and the spatial pattern of the loss of forests that occurred between 2002 and 2016 are shown in Fig. 1.

Land cover maps and variables for change analysis

Thematic land cover maps used in this research were produced by the Colombian Amazon Land Cover Monitoring System (SIMCOBA) of the Amazon Institute for Scientific Research SINCHI (https://siatac.co/simcoba/). SIMCOBA has prepared land cover maps for the periods 2002, 2007, 2012, 2014, 2016 and 2018. Three of the land cover maps prepared were used in this study: 2002, 2016 and 2018 a scale of 1:100,000³³. The maps were generated from the visual interpretation of a mosaic of Landsat 5 Thematic Mapper (TM) and Landsat 8 Operational Land Imager (OLI) images, using the PIAO technique (Photo Interprétation Assistée par Ordinateur). The classification categories of the land cover maps were based on the Corine land cover methodology adapted for Colombia³⁷.

The SIMCOBA system calculates the annual rates of Amazon forest loss (forest loss/ha/annual) by comparing the cover maps of the last two periods and subtracting from the previous map those forests that are no longer present in the most current map (Fig. 3). This process only considers the forests loss and the permanent forests. New forests due to natural regeneration or restoration are omitted in the calculations⁶.

To facilitate the interpretation of changes and cover transitions, the classification categories of the maps were re-categorized into 7 types: “Amazon forests”, “floodplain forests”, “fragmented forests and secondary vegetation”, “grasslands and shrublands”, “water bodies and wetlands”, “pastures and crops” and “urban and artificialized cover”. The land cover maps were resampled at a resolution of 60 m × 60 m to facilitate the computational analysis of the explanatory model, the simulations of the scenarios, and to keep the detailed spatial resolution of the coverage and explanatory variables¹⁶.

A geospatial database was created with a set of variables for the cover changes to create an explanatory model for each transition. Driving factors of change are grouped into the following variables: (1) accessibility, (2) climate, (3) landscape features, (4) production practices and environmental degradation, (5) landscape management, (6) socioeconomy, and (7) soil characteristics. We considered 41 explanatory variables (see supplementary information Table S1).

Accessibility variables such as roads and navigable rivers were obtained from the geodatabase at a scale of 1:100,000 of the Agustín Codazzi Geographical Institute of Colombia (IGAC). Bioclimatic temperature data were obtained from Worldclim v1.4³⁸. Cover variables (e.g., patch sizes Amazon forests and distance to pastures and crops) were created using the software ArcGis (v.10.7.1)³⁹ from the 2002 land cover map to understand which drivers were more influential in the dynamics of land-use changes since 2002 that resulted in the distribution of land cover in 2016.

Degradation variables, such as advances of the agricultural frontier, were obtained from the Territorial Environmental Information System of the Colombian Amazon (SIAT-AC)⁴⁰; livestock density data came from the Colombian Agricultural Institute (ICA); the fire density were processed from MODIS and VIIRS images (https://siatac.co/puntos-de-calor/); and the location of mining titles was obtained from the National Mining Agency.

The information on the landscape features and socioeconomic variables was obtained from different sources: (1) the limit of the protected natural areas was provided by the National System of Protected Areas (SINAP)⁴¹, (2) the Amazon Forest Reserve areas (Second Law of 1959) were obtained from the Ministry of Environment and Sustainable Development (MADS), (3) the location of the indigenous reservations was provided by the Ministry of the Interior, and (4) the limits of the areas of Indigenous Reservations and the legal status of the Amazonian region were obtained from the SINCHI cartographic database⁴⁰.

Socioeconomic information was spatialized from data from the National Administrative Department of Statistics (DANE). Soil-type data were obtained from IGAG, and topographic and altitudinal variables were derived from a DEM at 100 m resolution from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER V003) sensor⁴². All explanatory variables were resampled at a resolution of 60 m.

Patterns of land cover changes and transitions

The transformation patterns of territory are mainly defined by human intentions and the activities that these groups plan to develop after making the land cover changes, as well as the dynamics of vegetation regeneration⁴³. In this study, these changes in the study area were obtained and analyzed employing the Land Change Modeller (LCM) module of TerrSet³⁴ and using the land cover maps for 2002 and 2016 as input information (Fig. 2).

To represent dynamics and changes in the vegetation during the study period, a total of 14 transitions of greater importance in terms of area were considered (transitions with an area < 5000 ha were ignored) to reduce the complexity in the analysis of land use changes from the multiple possibilities of transitions that can be configured (Table 1). Three submodels of changes that grouped transitions were defined⁸: (1) submodel of degradation was defined by the changes that “Amazon forests” and “floodplain forests” may experience towards “fragmented forests and secondary vegetation ” and “grassland and shrubland”, either from selective logging, fire or small-scale activities such as the establishment of crops or pastures amid continuous blocks of intact forests; (2) Submodel of substitution included the transition from “Amazon forests”, “floodplain forests”, and “fragmented forests and secondary vegetation” to “pastures and crops”. These transitions are usually illegal as the replacement of Amazonian forests to extensive areas of pasture for livestock takes place mainly on lands that belong to the Colombian state, with no authorization of the country’s environmental authorities. (3) The regeneration submodel grouped the transitions that show recovery from a degraded cover to a forest cover or areas with abandoned “pastures and crops” towards “fragmented forests and secondary vegetation” in recovery.

The results showed that the substitution submodel was the most important pattern of transformation in the Colombian Amazon from 2002 to 2016, because the transition to “pastures and crops” was ~ 1.85 million hectares (Table 1). The submodel of degradation was dominated by the transition from “Amazon forests” and “floodplain forests” to “fragmented forests and secondary vegetation” with ~ 0.66 million hectares. In the submodel of regeneration, the most significant process occurred with “pastures and crops” where 0.38 million hectares transitioned to “fragmented forests and secondary vegetation”, giving rise to important areas of vegetation in recovery.

Neural network analysis and land use change factors

The artificial neural network multi-layer perceptron (ANN-MLP) is a multivariate statistical algorithm of machine learning widely used in the analysis of factors associated with changes in land use^44,45,46. MLP algorithm used in the LCM is an adaptation specially designed for the land change analysis³⁴. ANN-MLP is a suitable method of classification to solve non-linear relationships in complex data sets^{47,48,49,50,51} such as those in this study. ANN-MLP is made up of the elements shown in Eq. (1):

where, ({x}_{i}) are the input values or training data of the variables; ({w}_{i}) are the weights of the respective variables in the neural network; and n is the number of variables⁵². To solve the relationship between the explanatory variables and the response variable, the network adds one or more hidden layers of neurons connected by nodes (multi-layer) to find the most appropriate solution in the model (learning phase). The performance of the algorithm is controlled by two training parameters that are key for the application of an ANN-MLP (the learning rate and the “momentum”) because they control the speed and efficiency of the learning process⁴⁹. The values of (z) are multiplied by the transfer function or sigmoidal activation function (f), whose output values, (y), are the probability of change for each transition⁵¹. Under this criterion, the relationship of the factors for land use change for the 14 transitions were analyzed using ANN-MLP in TerrSet³⁴, starting with a reduced set of driving factors. To avoid including slightly relevant information during the learning phase of the neural network, Cramer’s V coefficient test was performed, which indicates the degree of association of each driving factor with the distribution of the land cover categories, that is, variables with V ≥ 0.15^34,53. Finally, it is necessary to declare whether the variables are dynamic or static. This is important for the projection of cover simulations, because dynamic variables are responsible for change over time, such as proximity to roads or previously deforested areas; and they can be recalculated at regular intervals to update the progress of transitions over time during the course of a simulation⁵².

The results were evaluated using two precision statistics: the “accuracy rate” (AR) and the “skill measure” (SM), using a subset of validation data (50% training/50% testing). The “accuracy rate” evaluates the ability of the algorithm to predict the correct classes of validation pixels after each iteration to train the network³⁴. AR values greater than 70% indicate that the submodel has good explanatory power. The “skill measure” indicates whether the prediction results are better than chance. SM values vary from + 1 (perfect prediction) to − 1 (worse than chance); a value of 0 indicates that the results are not better than chance³⁴. The analyses were developed in the submodule “transition potentials” of the Land Change Modeller (LCM) of TerrSet. For each transition, a transition potential map was produced, whose pixels reported continuous values from 0 (no probability of change) to 1 (high probability of change) and evaluation metrics.

Scenarios narratives and parameterization of simulations

Since the Peace Agreement, pastures and crops have been expanding at an accelerated rate in the forests of the study area¹⁰. The rapid loss of tropical rainforests is threatening the integrity of protected areas and connectivity in the Amazon and in other natural regions⁸. Based on this framework, LULC change scenarios have been developed to explore the effect of different biophysical and socioeconomic factors on the future of land use^54,55,56 in the Colombian Amazon for the year 2040. For this purpose, three storylines were prepared by a group of 6 experts with wide environmental and scientific knowledge of the region, using interviews³⁰. Expert interviews were orientated to briefly narrate events and contextualize factors that should be present and will drive the future of the Amazon in the next 20 years. Based on the interview results, three narratives were defined to build three scenarios: the trend scenario (BAU), the extractivist scenario, and the sustainable development scenario (see Table 2).

In the trend scenario simulation, the Markov chain model in TerrSet³⁴ was used to obtain the probability matrix, and it was projected to 2040. A Markov chain model is a random stochastic process that calculates the probability of land cover permanence and transition based on the analysis of historical changes, providing a framework for analyzing future land use demand⁵⁷. Markov probability matrix have been widely used for the analysis and modeling of changes in LULC¹⁶. Such a matrix shows the probability of a land use/cover change from one state to another taking place within a specified time period⁵⁸. An advantage of the Markov model is that the transition probability matrix can be modified by taking account of future land-use demands, based on current or future land-use management policies or on the narratives of local experts knowledgeable about environmental issues.

Two external models of probability of change were created from the Markov chain matrix of the trend scenarios: one for the extractivist scenario and one for the sustainable development scenario. The probability potential change values were repeatedly modified to accomplish the projected land use demands in each scenario. Based on the annual rates of Amazon forest loss calculated between the years 2002 and 2016, an average was estimated, and this value was used to calculate the increase and reduction in forest loss in the extractivist scenario and in the sustainable development scenario (Figs. 3, 4).

Because the narrative adds importance to the current conservation agreements that have been developed in the study area⁵⁹, a map was generated for rural associations that have a conservation and production program as an incentive in the transition from “pastures and crops” to “fragmented forests and secondary vegetation”, using this transition as an approximation of the effect on the recovery of vegetation in these areas. The parameterization in the extractivist scenario simulation included a year-to-year increase in the rate of forest loss up to 40% to 2040 as compared to the average rate between 2008 and 2016, and it included a delimitation of areas for mining titles as an incentive for the potential loss of Amazonian forests that may occur in this scenario. The scenario for sustainable development involved an immediate reduction in forest loss (− 80% of the average annual rate) and a gradual reduction year by year to − 99% to the year 2040, as compared to the average rate between 2002 and 2016. A total restriction of the loss of forests in protected areas, indigenous reservations, and on land slopes greater than 100% was also applied. Similarly, sustainable development actions would impact productive reforestation in prioritized restoration areas as well as properties that make up rural associations where conservation agreements are developed.

Simulation in LCM and validation

The ANN-MLP analysis evaluates the effect of different drivers on land cover change over a historical period, and it generates a map of suitability or probability of change that indicates where change will potentially occur. The LUCC model operates under the assumption that these drivers will continue to act in the future⁶⁰. The Markov chain model, on the other hand, calculates the amount of change based on historical land use change data or an external model that reflects expectations under a particular scenario. Therefore, before projecting the simulation of scenarios, it is necessary to evaluate the precision of the model in order to simulate changes in land/use cover⁵¹. A spatial simulation was projected for 2018 and was compared with the land cover map for 2018 made by SINCHI Institute.

The precision of the model was evaluated based on two indices: (i) the general kappa index (K)^51,52 and (ii) the k index of agreement (KIA)⁶¹. These two indices are useful to evaluate the following: (1) the ability of the explanatory model (ANN-MLP) to predict the location of the changes and (2) the projection of the amounts of change of each land cover by Markov chain analysis.

The K index assesses the concordance of the simulation in general for all covers with respect to the real data provided by the land cover map for 2018, expressed in values of 0–1; a kappa greater than 0.80 is a reasonable level of agreement. The KIA for land cover is a statistical measure of the difference between an observed agreement between two classifications versus agreement by chance⁶¹. KIA goes from 0–1; values of 0 mean no better than chance, and 1 means a perfect match.

Source: Ecology - nature.com