Abstract

Advances in computer vision are improving the ability to accurately extract structured information from frequent and high-resolution satellite imagery, shedding light on global challenges and furthering Sustainable Development Goals. While these advances, along with increased availability of high capacity computational resources, result in improved models, lack of diverse training data significantly limits applications of these models to certain geographical regions. We review state-of-the-art models for road detection using satellite imagery, and compare predictions of two models (one trained in Las Vegas, USA and another in Khartoum, Sudan) in Khartoum. This comparison shows the need for regionally trained models using local training data. Finally, we outline a roadmap to use transfer learning and regional models in cities that do not have human verified labels.

“Detecting Roads from Satellite Imagery in the Developing World”: abstract, paper, blog post

Presented at CVPR 2019 Workshop on Computer Vision for Global Challenges

Figure 1 — Prediction results in Khartoum for three different scenes. Top: input imagery. Middle: prediction results (shaded) from the model trained in Las Vegas overlaid with labels (red lines) on top of the input imagery. Bottom: prediction results
Figure 2 — (a) Labels generated by experts for validation. (b) Predictions from the Khartoum Model. (c) Predictions from Kumasi Model. (d) Predictions from Khartoum Model retrained in Kumasi with 10K steps.
Table 1 — Accuracy results for each model presented for Average (all classes), Road class and Background class.
Previous

Generating a Training Dataset for Land Cover Classification to Advance Global Development
Next

Deep Transfer Learning for Land Cover Classification on Open Multispectral Satellite Imagery