Land Boundary Delineation using U-Nets
summary
A segmentation pipeline uses U-Net with ResNet34 encoder pretrained on ImageNet to learn parcel boundaries from Sentinel-2 RGB (B2,B3,B4) tiles (2017), with augmentation (horizontal/vertical flips, random and center crops). The model outputs binary boundary masks and achieves intersection-over-union ~0.87 and recall ~0.83 on preliminary tests, so it works reasonably well given limited data; training is slow on CPU and future plans include larger time-series data and trying DenseNet-style encoders.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:20About my company
- 1:54Impact of the project
- 5:01Objective of the project
- 6:10Data Source
- 10:02Sample Image
- 13:07Preliminary Results
- 14:23Encouragement
- 16:46Future Work
- 19:03Limitations
- 20:19References
related talks
41mLarge Scale Field Delineation
36mClassification of Satellite Imagery
33mWhat and where? - Machine learning for geospatial image analysis
27mU-Net is all you need: applying the same AI approach for many different business domains
01h 36mFrom Maps to Models: An End-to-End Journey in Geospatial Machine Learning with Python
36mDeep Learning - Segmentation, Identification and Classification
45mTech Tutorial: Image segmentation
31mHow our neural networks find polygons to understand your documents better