segformer_b1-nlver_finetuned-1024-1024

NL-aerial segmentation

This is a segformer segmentation model finetuned on random samples from the entire 41,000 square kilometer of aerial photography data, see pdok aerial data, and using the BGT, see pdok BGT data. Specifically it takes in 1024x1024 aerial photographs taken at a resolution of 8cm/pixel and predicts water, buildings, roads/pavement, vegetation. This model is part of the NL-veranderdetectie project, summary in dutch. The model was trained using this codebase, see the repo for more details.

Model Details

Regular segformer, with the classification head adjusted to 5 classes.

Model Description

• Developed by: Het Waterschapshuis, in collaboration with the Dutch Waterboards [More Information Needed]

<!-- - **Funded by [optional]:** [] --> <!-- - **Shared by [optional]:** [More Information Needed] -->

• Model type: [Segformer Semantic Segmentation Model]
• License: [MIT]
• Finetuned from model: [nvidia/segformer-b1-finetuned-cityscapes-1024-1024]

Model Sources [optional]

<!-- Provide the basic links for the model. -->

• Repository: https://gitlab.com/hetwaterschapshuis/kenniscentrum/tooling/nlveranderdetectie/

<!-- - **Paper [optional]:** [More Information Needed] - **Demo [optional]:** [More Information Needed] -->

Uses

This model is used to create a segmentation map of the dutch waterways based on aerial imagery. This segmentation map is then used to actualize the BGT, a national map maintained by the Dutch government.

Training Hyperparameters

See the nlveranderdetectie repostiory for details of the training setup. https://gitlab.com/hetwaterschapshuis/kenniscentrum/tooling/nlveranderdetectie/ , specifically look at the yaml configs in the run/configs directory.

Hardware

The model was trained on a single nvidia 3090 GPU for ~2 days or ~126k forward passes with a batch size of 8.

Model Card Authors

The model car was written by J. Gerbscheid, email: [email protected]

Model Card Contact

The model was trained by J. Gerbscheid, email: [email protected]