Geo-referencing crop labels from street-level images using Structure from Motion
Agriculture and fishing
2022
Ground-truth labels on crop locations are critically needed to develop machine learning models to accurately identify croplands on satellite imagery for rural planning and combating food insecurity. Generating large quantities of crop labels from the field is challenging and financially prohibitive, especially in Sub-Saharan Africa, where most fields are small and difficult to access. Previous work proposed Street2Sat, a framework to tackle this by automatically generating crop-type labels from street-level images taken with vehicle-mounted cameras. Their framework geo-referenced the imaged crops using camera ratios to estimate the distance of crop fields from the camera location, which requires assumptions about the true height of the observed crop. This paper proposes an alternative method of geo-referencing crop type labels in Street2Sat by estimating the distance from the camera to roadside crops in Uganda using Structure from Motion (SfM), which matches features detected in consecutive overlapping image pairs and estimates the distance to the features using triangulation. We also evaluated how the distance estimation accuracy was influenced by feature extraction and matching algorithms and the presence of distinguishable features in the overlapping images.
Go to the output
Research Type
Technology development and applications
Organisation(s)
University of Maryland, National University of Singapore, École Centrale de Nantes, Arizona State University
Authors
Catherine L Nakalembe, Hannah R Kerner, Sneha Manimurugan, Ravi Singaram