What and where? - Machine learning for geospatial image analysis
summary
Project uses convolutional neural networks for per-pixel segmentation (based on SegNet) implemented in TensorFlow on 1 m-resolution GeoTIFF aerial images of Norway, training automatically from existing building vectors. Multispectral (near‑infrared) input improves accuracy but outputs are “blobby” shapes, so a post‑processing pipeline (mapAID) vectorizes to GeoJSON, straightens edges and highlights model failure cases for targeted retraining. Proof‑of‑concept finds many unregistered buildings; plan is to fine‑tune with local Tanzanian labels and move to DeepLab v3+ for better boundaries.
AI-written summary · about
chapters (12)
- 0:00The problem of no maps
- 1:51Mapping for safety in Tanzania
- 4:03Introduction to automatic mapping
- 7:22Building detection challenges
- 9:16Understanding neural networks
- 14:07Training the AI model
- 17:28Image segmentation techniques
- 20:28Results and data limitations
- 24:27Applications and illegal buildings
- 26:44Future improvements in mapping
- 29:14Post-processing and conclusion
- 31:44Q&A and closing
related talks
36mClassification of Satellite Imagery
31mFrom Radar Echoes To Real-Time Recognition: Deep Learning on Another Planet -An Open Source Workflow
17mGeoAI: a feature extraction tool set using Tensor Flow and Keras...
36mDeep Learning - Segmentation, Identification and Classification
01h 36mFrom Maps to Models: An End-to-End Journey in Geospatial Machine Learning with Python
28mUsing big data and AI to boost geospatial insights for hazard prediction
28mGeoAI: Evolution of AI-Based Tree/Vegetation Detection Models Using Imagery
45mTech Tutorial: Image segmentation