Mapping Clandestine Roads in Central America’s Protected Areas Affected by Narco-trafficking Using Satellite Imagery and Machine Learning

Keywords: informal roads, illicit infrastructure, Central America, deep learning, indirect land change
Abstract Type: Paper Abstract

Abstract

This study aims to assess the impact of narco-trafficking activities on land cover and land use change in Central America. These illicit activities target frontier forests, making detection and enforcement challenging. The study focuses on monitoring the development of clandestine infrastructure such as informal roads and airstrips, which are likely indicators of narco-trafficking activity. Multi-source satellite datasets from 2000 to 2022 and machine learning computer vision algorithms, such as deep convolutional networks, are utilized to detect clandestine infrastructure in the region. However, narrow roads and airstrips are difficult to identify using medium resolution satellite sensors. To address this, 4.77-meter high-resolution PlanetScope basemaps provided by Norway's International Climate and Forests Initiative (NICFI) were hand-labeled to identify informal roads and airstrips in three protected areas in Central America: Osa Peninsula (Costa Rica), Rio Platano Biosphere Reserve (Honduras), and Maya Biosphere Reserve (Guatemala). Since high-resolution PlanetScope data is only available from 2015, publicly available 30-meter Landsat, 25-meter ALOS PALSAR, and commercial 5-meter RapidEye satellite datasets were used to cover the earlier years. A machine learning model was trained on these sensors, using the 4.77-meter Planet labels as reference, to estimate roads and airstrips. The goal of this study is to build a time-series road map from 2000 to 2022 in Osa Peninsula using the trained model, allowing us to identify patterns and hotspots of clandestine infrastructure growth.