OpenEarthMap

The project uses the freely available satellite imagery and corresponding masks from the OpenEarthMap project to train an U-Net with EfficenctNet B4 backbone.

Satellite imagery and masks

The satellite imagery and the corresponding masks are not available on this github repository and have to be obtained from OpenEarthMap.

The model

The model was trained in Keras and most of the hyperparameters are used as suggested by Xia et al. 2022. The model's performance is shown in the tables bellow and correspond to the one found in the original paper. The pre-trained U-Net model with EfficenctNet B4 backbone is freely available and can be used for inference on satellite images of size 512x512 and a resolution between 0.5m-1m per pixel . Higher resolution usually results in better performance. The code for model training and evaluation can be found in .

Without TTA

Metric	Bareland	Rangeland	Developed Space	Road	Tree	Water	Agriculture	Building	Avg.
IoU	53.70	53.82	56.21	62.09	69.44	85.49	77.74	79.85	67.29
F1 score	0.70	0.70	0.72	0.77	0.82	0.92	0.87	0.89	0.80

With TTA

Metric	Bareland	Rangeland	Developed Space	Road	Tree	Water	Agriculture	Building	Avg.
IoU	57.96	56.13	58.09	64.85	70.57	86.04	79.82	80.92	69.30
F1 score	0.730	0.719	0.735	0.787	0.827	0.925	0.887	0.895	0.813

With TTA only in developed countries

Metric	Bareland	Rangeland	Developed Space	Road	Tree	Water	Agriculture	Building	Avg.
IoU	45.59	54.96	55.19	64.72	73.73	88.62	83.42	80.54	68.31
F1 score	0.604	0.709	0.711	0.786	0.847	0.940	0.909	0.892	0.800