Illegal land activities account for more than 40% of deforestation worldwide. Earth Observation technologies offer the capacity to objectively monitor the land-use changes associated with these important activities, but it is difficult to clearly identify and characterize specific land-use changes as “illicit.” I tackle this problem by building hypotheses based on observed illicit land patterns and then differentiating them through state-of-the-art algorithms over satellite imagery. I first hypothesize that unique, observable spatial patterns can be linked to known historical and institutional processes that drive them. Second, I leverage known observations of illicit activities to classify pixel patterns or objects that exhibit statistically similar variables using a deep learning (DL) algorithm. Using the Colombian Amazon region as a hotspot of biodiversity but also affected by illicit land activities, I detect a recent, explosive conversion of forests to cattle ranching outside the agricultural frontier and within protected areas since the negotiation phase of the Colombian peace accord. In contrast, coca farming is remarkably persistent across time, in which crop substitution programs remain ineffective to stop the expansion of coca farming deeper into Protected Areas such as Macarena, Nukak-Maku, La Paya, and Tinigua.
My conceptual contribution indicates key insights. First, countering common narratives, there is very little evidence that coca farming precedes cattle ranching. Second, spatiotemporal dynamics of illicit activities reflect how land agrarian policies encourage people to ranch cattle showing a linear increase during conflict (1985-2011) but an acute increase during post-conflict (2012-2019). Third, Colombia’s war on drugs (i.e., aerial fumigation and substitution programs) attenuated coca within the legal frontier, but expansion now is deeper into hotspots of biodiversity. The peace accord motivates different actors to expand illicit land activities at a rate never seen before. Finally, the theoretical framework allows linkage between patterns of illicit activities to build hypotheses that are then corroborated by deep learning models. This framework can be extended to many other issues associated with illicitness, such as illegal fishing, illegal mining, illegal commercial agriculture (e.g., soy, oil palm), and illegal lumber harvesting.
coca, cattle ranching, deep learning, Landsat, armed conflict, forest