Automatic Detection of Solar Panels on High Resolution Imagery.

Involved Models

Dataset are made with high-resolution satellite imagery from Nearmap (https://www.nearmap.com/au/en)
Labelme (https://github.com/wkentaro/labelme) is used as the tool to label the images.
Training dataset contains 3936 256x256 rgb images and labels which are collected from Capalaba, Springfield, New Farm, Fairfield, Sunnybank Hills in Brisbane, Australia.
Validation dataset contains 1344 images and labels which are collected from Springfield and Sunnybank Hills.
Test dataset contains 1360 images and labels which are collected from a suburb in Perth, Australia.

SegNet 0: SegNet with 5 encoders and 5 decoders (Original SegNet).
SegNet 1: SegNet with 4 encoders and 4 decoders.
SegNet 2: SegNet with 5 encoders and 5 decoders, each encoder is replaced by a ResNet block (Block with 3 convolutional layers).
SegNet 3: SegNet with 5 encoders and 5 decoders, each encoder is replaced by a ResNet block (Block with dynamic number of convolutional layers).
Fast SCNN 0: Original Fast SCNN.
Fast SCNN 1: Fast SCNN with the first two DSConv layers removed, modified Upsample layers.
Fast SCNN 2: Fast SCNN with the first two DSConv layers replaced with Conv layers, modified Upsample layers.

Name		Name	Last commit message	Last commit date
Latest commit History 11 Commits
data_processing		data_processing
metrics		metrics
model_application		model_application
pooling_layers		pooling_layers
trained_models		trained_models
training_logs		training_logs
evaluate_all_models.ipynb		evaluate_all_models.ipynb
fast_scnn_original.ipynb		fast_scnn_original.ipynb
fast_scnn_v1.ipynb		fast_scnn_v1.ipynb
fast_scnn_v2.ipynb		fast_scnn_v2.ipynb
readme.md		readme.md
requirements.txt		requirements.txt
segnet_4_encoder_decoder_v1.ipynb		segnet_4_encoder_decoder_v1.ipynb
segnet_original.ipynb		segnet_original.ipynb
segnet_resnet_v1.ipynb		segnet_resnet_v1.ipynb
segnet_resnet_v2.ipynb		segnet_resnet_v2.ipynb