Detecting Road Conditions from Space Using PyTorch, Public Data, and Free Satellite Imagery
summary
Fine-tunes a ResNet-50 PyTorch model on a Geolab satellite subset and validates against VDOT pavement labels, training on ~2,136 images using Google Colab TPUs and class-balanced sampling. Collapses VDOT’s five labels to three, hits about 66% accuracy, with poor roads identified well (≈70% recall) but fair vs good commonly confused; training overfits after ~4 epochs. Work uses low-resolution screenshots and limited data so it’s proof-of-concept rather than production-ready.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 3:55Road condition problem
- 5:24Transfer learning with ResNet50
- 6:29Data challenges
- 9:23Overfitting and limited data
- 10:26Free satellite imagery
- 11:31Geospatial data with GeoPandas
- 13:20Simplifying categories
- 15:06Predicting with PyTorch
- 17:12Model evaluation
- 20:12Confusion matrix and report
- 23:27Takeaways and limitations
- 25:15Future resources
- 25:58Question and answer session
related talks
06mAssisted Road Surface Tagging using Machine Learning
36mClassification of Satellite Imagery
29mRecognizing Traffic Lights With Deep Learning
46mDeep Learning for Satellite Image Processing, Lessons Learned
24mNatural Disaster Damage Assessment from Satellite Imagery
33mWhat and where? - Machine learning for geospatial image analysis
01h 13mExperimental Machine Learning with Holoviz and PyTorch in Jupeyterlab
18mDeep learning for Space imagery applications