A spatio-temporal Bayesian Network approach for deforestation prediction in an Amazon rainforest expansion frontier

Abstract

In the last decade, Brazil has successfully managed to reduce deforestation in the Amazon forest. However, continued increases in annual deforestation rates call for environmental modeling to support short-term decision-making. This paper presents the functioning of a stepwise spatio-temporal Bayesian Network approach for spatially explicit analysis of deforestation risk based on observation data. The study area comprises a deforestation expansion frontier located in the southwest of Pará state, Brazil. The proposed approach has been successful in estimating deforestation risk over the years. Among the selected variables to compose the Bayesian Network model, distance from hot spots and distance from degraded areas present the highest contribution, while protected areas variable present a significant mitigation effect on the phenomenon. Accuracy assessment indices corroborate the agreement between deforestation events and predictions.

Introduction

There has been great concern regarding deforestation in tropical regions since rainforests play a crucial role in the global climate system and in the carbon balance (Lawrence and Vandecar, 2014). Considerable efforts have been made to estimate, monitor, and understand deforestation patterns and drivers in Southern Asia (Kamlun et al., 2016, Reiche et al., 2015), in Central-South Africa (Grinand et al., 2013, Tegegne et al., 2016) and, particularly, in the Amazon region in South America (Barber et al., 2014, Barona et al., 2010, Sales et al., 2017, Souza et al., 2013). Since the 1980s, the Brazilian National Institute for Space Research (INPE) has been monitoring deforestation in the Brazilian Amazon forest under the project: Brazilian Amazon forest monitoring by satellite (PRODES), which carries out an annual mapping of clear-cutting areas, i.e., areas where the forest cover has been completely removed (Câmara et al., 2013, INPE, 2017a). PRODES project data are the basis for the Brazilian government to establish public policies in order to reduce deforestation (Arima et al., 2014). Despite being a successful example in reducing deforestation, Brazil needs to decrease even more its annual rates to conform to the National Climate Change Plan (Arima et al., 2014). However, PRODES project has reported an increase of annual deforestation rates after 2012 when the lowest rate was reached (INPE, 2017a).

Although livestock and agriculture expansion are significant deforestation drivers (Barona et al., 2010, Soler et al., 2014), there are previous factors that should be taken into account in land cover change modeling. For example, fires resulting from human activities are strongly associated with the encroachment of farming and logging activities into forests (Setzer et al., 2012, Tasker and Arima, 2016). Usually, forests are degraded by selective logging before being entirely deforested and converted to pasture/cropland in the following years (Pinheiro et al., 2016). In this regard, INPE’s Amazon monitoring program also counts on other systems besides PRODES project such as DEGRAD,¹ which maps degraded forested areas, i.e., forest cover areas that have not yet been totally removed; and Wildfire system, which detects vegetation fires in satellite images (INPE, 2008). Predicting land demand and allocation pushed by the deforestation process is a challenging task (Dalla-Nora et al., 2014, Rosa et al., 2014). In order to do so, a diversity of spatially explicit and predictive land-use modeling approaches such as Logistic Regression (Rosa et al., 2015, Rosa et al., 2013); Maximum Entropy (de Souza and De Marco, 2014); Cluster-Analysis (Schielein and Börner, 2018); Cellular Automata (Roriz et al., 2017, Yanai et al., 2012); Genetic Algorithm (Soares-Filho et al., 2013); Geostatistical Hurdle Model (Sales et al., 2017); Ensemble Model combining Random Forest, Logistic Regression and Neural Network (Bradley et al., 2017); among others, have been proposed to identify regions where changes are most likely to occur. Indeed, there are different land use modeling approaches in the literature, and several reviews have classified them in different ways (Brown et al., 2013, Chang-Martínez et al., 2015, Noszczyk, 2018, Rosa et al., 2014). However, none of these reviews mention the Bayesian Networks (BNs) as an approach for land use modeling, although they offer advantages compared to other approaches and have the potential to deal with spatial information (Aguilera et al., 2011, Landuyt et al., 2013, Uusitalo, 2007).

On the other hand, various authors have employed BNs for land use modeling. Celio et al. (2014), for instance, applied BNs to model land use decisions and analyze the driving forces of changes. Krüger and Lakes (2015) used BNs to address and quantify the uncertainty in land change modeling. Dlamini (2016) proposed a BNs approach to identify high deforestation risk areas and its underlying causes, while Mayfield et al. (2017) used several statistical and machine learning techniques, including BNs, to predict deforestation. In turn, Li et al. (2016) used BNs to infer about urban land use. Additionally, BNs have increasingly been used as decision support tools for environmental management. For example, Dlamini (2010) presented a BN approach for estimating the likelihood of wildfire occurrence, while McCloskey et al. (2011) proposed an integrated methodology to identify suitable areas for urban development and conservation units. Pérez-Miñana et al. (2012) used BNs to estimate greenhouse gas emissions in the agricultural sector, and Madadgar and Moradkhani (2014) proposed a spatio-temporal BN model to predict droughts. In turn, Lucena-Moya et al. (2015) presented a BN model to predict the impacts of climate change on ecosystems. Gonzalez-Redin et al. (2016) presented a methodology to analyze the trade-offs between timber production and biodiversity conservation using BNs, while Silva et al. (2017) employed BNs to predict areas for sugarcane expansion.

BNs can be classified as a Machine Learning and Statistical Model. According to the description of Brown et al. (2013) and Chang-Martínez et al. (2015), this type of model relies on algorithms and rigorous statistical methods to encode relationships between land use changes and drivers, and aims to estimate the quantity and/or spatial allocation of land use changes at some point in the future. Although land use models are useful for representing scenarios likely to change in the future (Dalla-Nora et al., 2014), predictions tend to become increasingly uncertain over time (Rosa et al., 2014). Regarding deforestation predictions, this uncertainty is due to a large number of economic, political, and social variables that directly or indirectly influence the deforestation process and cannot be accurately accounted for (Dalla-Nora et al., 2014, Rosa et al., 2014). In this sense, short-term prediction models are more advisable, mainly when using variables that represent human behavior as indicators of deforestation, since human behavior constantly changes and is therefore hard to predict.

Employing BNs as a modeling approach to predict deforestation risk is not unheard of (Dlamini, 2016, Krüger and Lakes, 2015, Mayfield et al., 2017). However, temporal domain has never been taken into account and deforestation has been considered as a static process when modeled by BNs. In this context, the spatio-temporal BN approach proposed in this work is a stepwise application of a BN model over time, as an alternative to dealing with deforestation process temporal dynamics in BN modeling, which apparently, has not yet been performed. For that reason, we employed the e-BayNeRD method (Silva et al., 2017) – enhanced Bayesian Network for Raster Data — to predict deforestation risk at an expansion frontier located in the center of the Brazilian Legal Amazon. Therefore, the objectives of this study are (i) to explore the capability of the proposed approach for short-term predictions of areas susceptible to deforestation and (ii) to verify the importance of the chosen variables as deforestation risk indicators.