Deep Learning based road extraction from historical maps

This repository contains the code, test patches and weights for the paper [Deep Learning based road extraction from historical maps]

Dataset (Source)

---

Turkey 1:200k Historical Topographic Maps

The historical DHK 200 Turkey map used in this study covers a large area of around 150,000 square km in northwest Turkey, including the regions of Ankara and Bursa. The DHK 200 Turkey map legends are organized bilingually in accordance with the rest of the World War II German military maps [1].

Model	Batch-Size	F-1 Score	No. Of Params	Weights
Timm-resnest200e(U-Net++)	16 Batch-Size	0,577	68M	Timm-resnest200e.pth
Timm-resnest200e(U-Net++ scSE)	16 Batch-Size	0,542	68M	scSE timm-resnest200e.pth
Timm-resnest200e(MA-Net)	16 Batch-Size	0,541	68M	MA-Net timm-resnest200e.pth
Inceptionv4(U-Net++)	16 Batch-Size	0,525	54M	Inceptionv4.pth
Densenet201(U-Net++)	16 Batch-Size	0,511	18M	Densenet201.pth
Resnext50_32x4d(U-Net++)	16 Batch-Size	0,491	42M	Resnext50_32x4d.pth

Model	Batch-Size	F-1 Score	No. Of Params	Weights
Timm-resnest200e(U-Net++)	8 Batch-Size	0,564	68M	Timm-resnest200e.pth
inceptionresnetv2 (U-Net++)	8 Batch-Size	0,501	54M	inceptionresnetv2.pth
densenet201(U-Net++)	8 Batch-Size	0,485	18M	densenet201.pth
resnext50_32x4d(U-Net++)	8 Batch-Size	0,472	42M	resnext50_32x4d.pth
efficientnet-b1(U-Net++)	8 Batch-Size	0,4542	6M	efficientnet-b1

Framework

---

Outputs

---

Classwise Metrics for best Models

1	2	3
4	5	6

Prequsities

The code was implemented in Python(3.8) and PyTroch(1.14.0) on Windows OS. The Qubvel segmentation models pytorch library is used as a baseline for implementation. Apart from main data science libraries, RS-specific libraries such as GDAL, rasterio, and tifffile are also required.

Citation

Avcı C., Sertel E. , Kabadayı M. E. “Deep Learning Based Road Extraction from Historical Maps”, IEEE GEOSCIENCE AND REMOTE SENSING LETTERS,Accepted.

Contact Information:

Cengiz Avcı - avcice16@itu.edu.tr