Effects of armed conflict on deforestation and coca crops

This study aims to identify, spatialise and quantify coca crops and deforestation in an emblematic Natural National Park in Colombia (“Serrania La Macarena”, The Site). As consequences of the intensity of armed conflict in a temporal scope of eight years (four years before and after the peace agreement in 2016), applying a supervised random forest technique.

The Site is located in the centre of Colombia with a complex geological configuration and archaeological, environmental and ethnic importance. Its location is strategic for connecting ecosystems across the country in the northern and the Amazonian forest in the south. Its extension is approximately 6000 km2 with an average temperature of 27°C and topography between 200 to 400 height above sea level. This national park is one of Colombia's most representative national parks due to an exceptional place called “Caño Cristales”, the most important tourist place in this park (CNPPA, 1982; PNN of Colombia, 2018). This river has an endemic aquatic plant (Macarenia clavigera) that can vary from blue, red, yellow to green, generating a remarkable ecotourist opportunity. Nonetheless, security issues caused by illegal groups like FARC reduce this possibility in the long term. 

Five NDVI maps were obtained, evidencing significant changes across this temporal scope. The main changes are in the western and southern boundary of the National Park, with the lowest values in 2014 and subsequently in 2018. Besides, there is a geological configuration in the southwest part of the park, always with medium NDVI values. See NDVI changes across time in the maps below (Figure 1).

The samples amount to conduct the supervised classification were between 140 to 181. The lowest amount of samples were in 2012 due to image limitations (see limitation section below). The number of bare soil samples in 2014 and 2018 were higher since burned areas were easy to identify during those years. Finally, I tried to keep the same proportion of samples for the remaining land cover units. See the number of samples per land cover per year below.

The accuracy assessment and Kappa statistics were greater for all models with NDVI as a variable. Kappa values were between 87.41% to 92.09%, and accuracy assessment between 94.03% and 95.81%. The lowest values for both assessments were in 2012. See consolidated information in Table 2.  Therefore, the model with bands between 2 to 7 and NDVI values was selected to identify coca crops and forest loss mainly, using the random forest technique.

The importance per variable was calculated, identifying that in 2012 and 2014, B5 was the variables with more significant permuted importance, followed by B7 and NDVI in 2012 and B6 and NDVI in 2014. The permuted importance behaviour is quite similar in 2016, 2018 and 2020, having B6 as the most important variable, followed by B7 and NDVI variable mainly, and finally, Band 4 was the variable with the lowest importance.

Five land cover units were identified in the Site, but it is mainly covered by forest. Nonetheless, there is a remarkable pattern that corresponds to the land cover change from forest to grass and subsequently to coca crops. Besides, there are changes in the centre of the park from north to south where crosses the "Cano cristales" stream, which is adjacent to the bare soil in the southwestern part of the park. See maps below.

The forest loss between 2012 to 2016 was 40.28 km2 (6.92km2 per year), and the coca crops increment was 28.63 km2. By contrast, after the peace agreement, the forest loss rate did not vary significantly, and between 2018 and 2020, the forest cover increased by 12.24 km2. Although between 2016 and 2018, the coca crops increased substantially (23.96 km2), these results represent a positive panorama for the coming years under post peace agreement since between 2018 and 2020, coca crops area decreased 23.52 km2. Therefore, the forest loss was more significant before 2016, with abrupt change across 2012 to 2016. Moreover, this rate is inversely proportional to the coca crops, increasing substantially between 2012 to 2018. In terms of grass area, from 2012 to 2016, the grass areas rose similarly to the coca crops, but after 2016, it reduced its extent. Finally, bare soils fluctuate across time without dramatic changes.

A confusion matrix for each dataset (training and testing data) per year was obtained, analysing each land cover with an approximate total of 160000 samples based on the total site area (6000 km2, approx).  See Table 3 to 7 confusion matrices for each year divided into testing and training data.

Kappa statistics and accuracy assessment were presented in Table 2, manifesting a remarkable performance of these models with values greater than 87%. Nonetheless, this study aims to identify deforestation and coca cultivation; thereby, analysing these statistics per land cover individually could provide quantifiable limitations to the target land cover units. The lowest accuracy assessment among land cover units were coca crops with values between 88 to 90%, followed by grassland cover. Besides, the forest land cover obtained an accuracy assessment between 94 to 96%. See accuracy assessment per land cover in Table 8. 

This study had demonstrated that before the peace agreement, the forest loss rate was increasing dramatically, and the coca crops and grassland. On the contrary, after the accord, those rates decreased substantially; these results oppose the findings of other studies, which argue that the peace agreement has been a driver of deforestation in Colombia (Clerici et al., 2020; Mendoza, 2020; Prem et al., 2020). The main reason for these findings could be associated with the significant increment of tourists in the national park due to the secure perception (PNN of Colombia, 2018), providing economic alternatives to small farmers inside the park rather than coca cultivation.

This supervised classification had better performance integrating NDVI as an analysis variable since it increases the permuted importance of red and near-infrared bands, enhancing the difference in wavelength between water and vegetation and bare soils. Besides, this study has shown that all bands integration helps identify diverse land covers since each band detects specific characteristics of each object expressed in the wavelength. Nonetheless, both models presented positive accuracy assessment and acceptable kappa statistics which can be applied on a large scale with a larger temporal scope. 

The first limitation was the scan line fault in Landsat 7 imagery for 2012 analysis since after conducting correction there was still an identifiable change in those areas (see map below). This constrain affects the supervised classification, over or sub estimating the land cover area. 

Secondly, the annual coca harvests in Colombia could be between 6 to 7 due to the perfect conditions for this plant development in Colombia (Angel, 2012). Therefore, I cannot identify specific patterns during the year since composite images are annual due to the availability of suitable satellite imagery for this analyses in shorter periods, generating limitations in terms of annual analysis. 

Finally, the supervised classification process misinterpreted some reduced areas of land covers due to the resolution of the Landsat imagery. For instance, small streams with longitude lower than 30 m were included in other land cover units (mainly forest) and young coca crops as grass and mature coca as forest. Therefore, field verification data is imperative to support the positive statistical results obtained in this study. Nonetheless, it is a good approach to prioritise specific areas inside extensive protected areas. 

This study has proved that the peace agreement has reduced the deforestation and coca crops in the “Sierra La Macarena” National Park, increasing the ecotourism business in the park and providing alternatives opportunities to local farmers whose main economic activity was illegal crops harvesting.

ANGEL, Y. 2012. Metodologia para identificar cultivos de coca mediante analysis de parametros red edge and espectrocopia de imagenes. Master of Geomatic National University of Colombia.

ARMENTERAS, D., CABRERA, E., RODRÍGUEZ, N. & RETANA, J. 2013. National and regional determinants of tropical deforestation in Colombia. Regional environmental change, 13, 1181-1193.

ARMENTERAS, D., SCHNEIDER, L. & DÁVALOS, L. M. 2019. Fires in protected areas reveal unforeseen costs of Colombian peace. Nature ecology & evolution, 3, 20-23.

BAGAN, H., MILLINGTON, A., TAKEUCHI, W. & YAMAGATA, Y. 2020. Spatiotemporal analysis of deforestation in the Chapare region of Bolivia using LANDSAT images. Land degradation & development, 31, 3024-3039.

CLERICI, N., ARMENTERAS, D., KAREIVA, P., BOTERO, R., RAMÍREZ-DELGADO, J. P., FORERO-MEDINA, G., OCHOA, J., PEDRAZA, C., SCHNEIDER, L., LORA, C., GÓMEZ, C., LINARES, M., HIRASHIKI, C. & BIGGS, D. 2020. Deforestation in Colombian protected areas increased during post-conflict periods. Scientific reports, 10, 4971-4971.

CNPPA 1982. IUCN Commission on National Parks and Protected Areas, IUCN, International Union for COnservation of Nature and Natural Resources.

MENDOZA, J. P. 2020. Colombia's transition to peace is enhancing coca-driven deforestation. Environmental research letters, 15, 1-9.

PENA, J. A., YUMIN, T., LIU, H., ZHAO, B., GARCIA, J. A. & PINTO, J. 2018. Remote sensing data fusion to detect Illicit crops and unauthorized airstrips. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives, 42, 1363-1368.

PESARESI, M. 2008. Textural analysis of coca plantations using 1-meter-resolution remotely-sensed data. In: CENTRE, E. C. J. R. (ed.). United Nations Office on Drugs and Crime (UNODC).

PNN OF COLOMBIA 2018. Parque National Natural Sierra de la Macarena. In: MACARENA, P. N. N. S. D. L. (ed.).

PREM, M., SAAVEDRA, S. & VARGAS, J. F. 2020. End-of-conflict deforestation: Evidence from Colombia’s peace agreement. World development, 129, 104852.

RUNAP. 2020. Total de hectareas en areas protegidas en Colombia [Online]. Available:  https://runap.parquesnacionales.gov.co/cifras  [Accessed].