Predicting Future Deforestation in the Congo Basin Using Machine Learning
Abstract
The purpose of this project is to produce a future deforestation risk map within the Democratic Republic of the Congo (DRC) and to make a free, open-source, and easy to use deforestation risk modeling pipeline so others can run their own deforestation predictions in their area of interest.
The DRC is home to the second largest tropical forest in the world and many critically endangered species. Understanding land change dynamics and predicting which areas are at higher risk for deforestation allows us to make informed conservation planning decisions, understand how food production impacts deforestation, and abate emissions from deforestation. We used the University of Maryland annual tree cover loss data, as well as deforestation driver datasets such as agricultural expansion, population, forest management activities, and access to forests, to develop a deforestation prediction model. We compared multiple tree-based machine learning algorithms and trained them by classifying annual tree cover loss data between 2012 and 2017. We then predicted future deforestation risk by feeding the model the latest versions of the predictor variables, producing a spatial map of both risk and likely future deforestation by the year 2025. The model results demonstrate that proximity to previous tree cover loss, distance to previous shifting cultivation, and distance to population centers among others have a significant influence on the risk of future deforestation. Results from the prediction model will be used by the World Resources Institute and the DRC government to understand the links between forest clearing and urban and peri-urban centers, which drive demand for charcoal, wood fuel, and food. Using agricultural expansion data, we ran different agricultural yield scenarios to understand how food production impacts deforestation in the region. Finally, we overlaid biomass data to run carbon emissions impacts and determine scenarios for the DRC in meeting its climate goals. The open-source model pipeline will include a user-interface where users can perform each step of the pipeline, from data pre-processing to model training and prediction, by providing simple text file inputs. The interface and the open-source code enables users of many backgrounds to do this type of deforestation risk analysis in any area of the world.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMIN51C..07L
- Keywords:
-
- 1640 Remote sensing;
- GLOBAL CHANGE;
- 1942 Machine learning;
- INFORMATICS;
- 4329 Sustainable development;
- NATURAL HAZARDS;
- 6620 Science policy;
- POLICY SCIENCES & PUBLIC ISSUES