The University of British Columbia
UBC - A Place of Mind
The University of British Columbia
Faculty of ForestryMGEM Project Library
  • Research Projects
    • Soil
    • 3D + GeoVis
    • Wildlife
    • Climate
    • Water
    • Urban
    • Wildfire
    • Forests
    • Carbon
  • Community Partners
  • Work With Us
  • Mentor Portal

Applying a Faster R-CNN Model Coupled with a Generalized Linear Model to Optical Imagery to Fill Population Data Gaps in Villages along the Southwest Coast of Madagascar

Abstract

Arid, infertile soil within the southwest of Madagascar creates poor agricultural conditions that have worsened over the years due to climate change. Coastal areas are experiencing rapid population growth due to the allure of marine resources for economic support and food security. However, there exists large uncertainty surrounding how many people rely on these marine resources. As a result, bottom-up management systems have been implemented to mitigate data paucity related to population and marine resource management. This study outlines a methodology for predicting population of coastal villages via detection of artisanal fishing boats, known as pirogues, and houses from the RGB bands in high resolution satellite imagery from Maxar Technologies using a Faster Regions with Convolutional Neural Networks (Faster R-CNN) detection model and a generalized linear model. The Faster R-CNN model trained for object detection had an initial accuracy of 87% for house detection and 75% for pirogue detection. Yet, this model’s accuracy varied from 0-11% for pirogues and 20-77% for houses when applied to three villages not seen previously. Misclassification was due to high reflectance from sand masking out boats and vegetation being improperly detected as houses and pirogues. Pirogue and house counts were found to be significant predictors of population within the generalized linear models created on 2015 census data. However, due to limited size and range of the training data the generalized linear model was not able to accurately predict 2021 population. Both the Faster R-CNN and generalized linear model show promise for the ability to predict population via high resolution (~24 cm) satellite imagery. Future methodology should include more bands than RGB if attainable, specifically those in the NIR range, and direct effort towards increasing the magnitude of training data to create more robust models before they can be confidently used.

MGEM Student: Olivia Waite
Community Partner: Blue Ventures

Explore this project

Download the final report and data

Cite this project

Waite, Olivia, 2022, “Applying a Faster R-CNN Model Coupled with a Generalized Linear Model to Optical Imagery to Fill Population Data Gaps in Villages along the Southwest Coast of Madagascar”, https://doi.org/10.5683/SP3/TNNBSF, Borealis, V1

Master of Geomatics for Environmental Management
Faculty of Forestry
University of British Columbia
2424 Main Mall
Vancouver, BC Canada V6T 1Z4
Back to top
The University of British Columbia
  • Emergency Procedures |
  • Terms of Use |
  • Copyright |
  • Accessibility