Towards Mapping Clandestine Infrastructure in Central Americas protected areas affected by narco-trafficking using Multi-sensor Data Fusion and Deep Learning
Abstract
Illicit activities such as narco-trafficking have large impacts on land cover and land use change in Central America. Such activities target frontier forests where detection and enforcement are difficult. Increases in clandestine transportation, via roads and airstrips, are crucial indicators of increased narco-trafficking activity in impacted regions. Roads and airstrips initially built by narcotraffickers may also be later used by others to invade previously inaccessible forested regions and harvest timber or clear land for cattle. Monitoring the development of illicit infrastructure annually in affected areas could improve estimates of both the direct and indirect narco-activity-related land cover and land use change, and the ability to monitor these changes. Currently, no annual maps of built infrastructure are available in protected areas of Central America. Multi-source satellite datasets and deep learning computer vision algorithms could be effective in extracting these built-up structures. We aim to generate annual infrastructure maps using optical and radar satellite imagery, such as Landsat, Sentinel-1 and 2, PALSAR, and high-resolution commercial Planetscope imagery from 2000-2020. Since clandestine infrastructure due to narco-trafficking are primarily small and medium roads and airstrips, Planetscopes commercial high spatial resolution optical imagery products will be crucial for accurate identification and labeling of infrastructure. We hand label Planetscope imagery across 1000 chips at 1024x1024 pixel size to train a convolutional network to detect roads, airstrips, and buildings using other public sensor inputs with different satellite combinations. Here we will show initial model results from using deep learning to identify clandestine infrastructure in 2017, comparing accuracy metrics (IoU) across 4 satellite sensors (Planetscope, Landsat, Sentinel-1, and Sentinel-2). The dataset and methods developed to automate road detection in Central Americas protected areas could be extended to identify informal road activity in protected areas elsewhere.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMGC44A..07M