satellite-image-recognition

satellite image recognition using U-Net model

demo

Data source

Get data from Google satellite map, and then labeled and splite. Then model is based on 54 (160px160px*3) RGB images data.

train

Training process is using reflection,mirror method and U-net model

RGB list

• [‘water’] = [48, 93, 254]
• [‘tree’] = [12, 169, 64]
• [‘playground’] = [139, 69, 19]
• [‘road’] = [47, 79, 79]
• [‘building_yard’] = [255, 255, 255]
• [‘bare_land’] = [239, 156, 119]
• [‘general_building’] = [249, 255, 25]
• [‘countryside’] = [227, 23, 33]
• [‘factory’] = [48, 254, 254]
• [‘shadow’] = [255, 1, 255]

quote from

Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 3431-3440.