Due to biogeographic conditions, Colombia has a great amount of biodiversity, but it is unfortunately among the top ten countries that have shown important loss of forest habitat. It has an annual destruction rate of 0.5% [1]. It is calculated that by the year 2050 almost 80% of species extinction will be the result of deforestation [2]. This is an ongoing cycle in Latin America that is brought about by the development of road infrastructure projects [3], which, without suitable environmental planning, contribute to the fragmentation of forests and biodiversity loss.

The main failures of environmental impact studies such as the descriptive nature of the assessments, the lack of measurable indicators and the absence of quantitative predictions [4] do not allow the potential impacts of new roads to be rigorously established, which could compromise the conservation of valuable resources in the long-term. These assessments should be complemented with spatial analysis models that help predict potential changes in the landscape caused by the construction of a road.

Statistical models have been widely used to model deforestation in the tropics, [5, 6] mainly in the Amazon [7,8]. Various pieces of research that included these models had identified the places most susceptible to deforestation and the main determinants. There has been no deep analysis or estimation of the condition of forests after the construction of new road infrastructure projects; this would help to improve the local management and conservation of forests.

Statistical models are attractive for decision makers due to their methodological stability and long application history. However, for these types of models, a priori assumptions should be made about the relationship between variables, which could not exist. The main advantage in the use of ANNs to make predictions is that this type of assumption is not required as they have the ability to capture nonlinear relationships. They can, therefore, be a more appropriate instrument when emphasis is placed on self prediction and not on the underlying relationships between independent and dependent variables [9].

ANNs have been widely applied in medicine and molecular biology. At the beginning of the 1990s the first studies were reported in the areas of ecology and environmental sciences [10]. Khoi and Murayama [11] used ANNs to predict the areas most vulnerable to deforestation, Larsen [12] used ANNs to evaluate the response of forest ecosystems in the presence of high amounts of carbon dioxide and ozone, and Pijanowski et al. [13] constructed an ANN model to predict changes in coverage and land use on a national scale.

The objective of this research is to construct neural network models, evaluate and choose the best model based on its performance (highest AUC value) to predict deforestation from 1980-2000, and evaluate the change in the probability of deforestation for a scenario that simulates the construction of new roads.

The networks that showed high AUC values were those trained with a lower training rate and greater amount of neurons in the hidden layer; this can be observed in Fig. 2.

The ANN that showed best predictive capacity obtained an AUC value of 0.8212 (Table 1). Based on this, a deforestation probability surface was generated for the area covered by forests in 1980 as well as the deforestation probability surface for the simulated scenario.

The comparison of the scenarios suggests that 1,152,565 pixels, approximately 103,729 ha of forest (pixel area of 900 m²), would increase its probability of deforestation (Fig. 3); 75,622ha showed a low increase (0-0.11), 20.186ha an average increase (0.11-0.37), and 7.922ha a high increase (0.37-0.9). However, some pixels, approximately 7%, showed a reduction in the probability of deforestation for the simulated scenario, which goes against what was expected based on the literature, which indicates that the probability of deforestation increases with the proximity to roads [34, 35]. This may be attributable to the expected error in the simulation, taking into account that the predictive capacity for the best ANN was approximately 82%.

Figure 2 Source: The Autors

Table 1

Source: The Autors

Figure 3 Source: The Autors

As the training rate decreased, the AUC value increased, but we must bear in mind that even though the networks with low training rates show less error, compared with other rates, training is slower. Therefore, the value of the learning rate should be large enough to allow a quick learning process but small enough to ensure their effectiveness. For the current research, the training parameters that showed the best performance in terms of training times and predictive capacity was 0.05 for training rate and 0.5 for the momentum.

Another alternative for training would be the use of an adaptive network with a resilient back propagation algorithm in which the rate changes automatically during the training. This means that the problem of defining a global training rate that is appropriate for the whole learning process [36] is avoided.

The models generated in the training phase did not show wide differences in terms of accuracy; the AUC values were between 0.79 and 0.82. This is consistent with Müller and Mburu [37] who also found similar results when modeling deforestation with different neural network architectures; and Okwuashi et al. [38] who simulated changes in land use in Lagos, Nigeria for three time periods, obtaining AUC values between 0.7574 and 0.7850.

The evaluation of the potential effect of the two roads proposed for the simulated scenario suggests an increase in the deforestation probability for 103.729ha of forest. Wilson et al. [39] in their study, carried out in the temperate zone forests in southern Chile, found an increase in deforestation probability of about 27,000ha after simulating the construction of a new road. In the central region of México, Nelson and Hellerstein [34] simulated the impact of new roads on deforestation and demonstrated that access to roads significantly influenced the change in land use.

The results of the research show an increase in the probability of deforestation within 6 km of roads; the probability diminishes as the distance increases. Mertens and Lambin [40] found a negative relationship between proximity to roads and deforestation at a distance beyond 7.5 km, and Barber et al. [36], in their study undertaken in the Amazon, found that 95% of the deforestation occurred within 5.5 km of roads. The results suggest that there is an agreement between different studies on the spatial pattern of environmental impacts resulting from the construction of roads. This offers the possibility to develop appropriate models for spatial scales that are larger than what is considered in this investigation.

Other than the immediate impact of road construction within or near to forest areas, roads are permanent elements of the landscape [41]. So, even when they do not directly influence deforestation, they represent a long-term determinant to forest conversion. This is the case with the simulation of the “El Bagre - San Jacinto del Cauca” road, which does not go through forest areas but generates an increase in the deforestation probability in the eastern part of the study area. This shows that the spatial configuration of the scenery is determined, to great extent, by the proximity of forests to roads.

Freitas et al. [41] indicate that although the variable of distance to roads is a strong predictor of forest dynamics in deforestation processes, the effect can only be detected when agricultural expansion has stabilized over a certain period. However, the effect of the construction of a new road such as the “variante Porce”, which would cross areas with low agricultural aptitude could be the increase in mining and excessive wood extraction and not the planting of agricultural crops. This is disturbing if we take into account that ecological studies suggest tropical forest based on poor soils have greater levels of biodiversity than those that have better quality soils [42].

The simulation shows the areas for which there is an increased deforestation probability are associated with the roads designed. Consequently, this estimation is a conservative one since many paths and temporal trails could be built from new roads facilitating access to forests. Therefore, the area that will experience an increase in deforestation could possibly be greater, which would put at risk the environmental services that house the forest ecosystems.

The simulation with the network that presented the best predictive capacity during the training process showed an increase in the deforestation probability for 103.729ha of forest. Although the results cannot be validated with real data since the proposed roads have not yet been built, the development of this type of analysis provides planners with an empirical orientation on the magnitude and location of damage. Therefore, after the completion of road projects strict surveillance by the environmental authorities is needed and the design of policies aimed at creating incentives for forest protection, which would enable the population to take part in forest conservation activities.

Although carrying out projections for change through empirical deforestation models generates uncertainty, the study shows how the ANN has the ability to capture changes and trends in terms of deforestation probability. Therefore, greater precision can be achieved by developing models for shorter time periods that help not to predict the exact location of future deforestation, but rather the places that have a high risk of conversion.

Furthermore, the ANNs could be important tools to select alternative routes of linear structures since they allow all the environmental information for the area in which it is intended for the civil work to be located to be integrated. This would help environmental science technicians to visualize landscape transformation while the location of the linear infrastructure is defined. This represents valuable information that is useful in order to formulate and implement compensation forestry projects under the mechanism for reducing emissions for deforestation and degradation (REDD).