Land delineation is a particular example of a segmentation problem. In this project we used U-Nets, a type of CNN, to learn the boundaries of parcels of lands from satellite images. The U-Nets model proved very efficient for the task.
2. Ishango.ai is a social enterprise that creates data science jobs in Africa.
It connects talents in Africa to global host companies to work on real-life data
science projects.
As a company, it has the mission to create high-skill data science jobs across
Africa.
3. Outline
1. Background
2. Objectives
3. Methodology
4. Sample Input Data
5. Architecture of U-net
6. Preliminary Results
7. Impact
8. Conclusion
9. Future Works
4. Background
As cities and towns develop, demand for land for various purposes such as
agriculture, building of homes, and industrialisation increases.
Leveraging on deep learning algorithms, land allocation and monitoring of land
use can become a very handy task.
In this project we are used U-net, a type Convolutional Neural Networks to learn
land boundaries from satellite images.
6. Objectives
● To develop a working model that can accurately find the boundaries of
parcels of land from satellite images.
● To apply the model in other areas such as agriculture and land registry.
7. Methodology
● Download satellite data (Images download period : 2017)
● Combine bands to form input RGB images
● Normalise mask layout of land and use as target
● Using U-Net;
○ Encoder : resnet34 and Encoder-weight: imagenet
○ Segment land boundaries from land
13. Future Works
● Further exploration can be done using other pre-trained models as encoder.
● Model could be trained on current dataset to learn new boundaries.
14. Limitation
● Long computation time during model training, GPUs are preferred
for this project.
● Downloaded image data requires high storage size.
15. References
● Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. "U-net: Convolutional
networks for biomedical image segmentation." International Conference on
Medical image computing and computer-assisted intervention. Springer,
Cham, 2015.
● Kumar, Sandeep, and Prabhu Jayagopal. "Delineation of field boundary from
multispectral satellite images through U-Net segmentation and template
matching." Ecological Informatics 64 (2021): 101370.
● https://smp.readthedocs.io/en/latest/quickstart.html