/Dubai-Satellite-Imagery-Multiclass-Segmentation

Simulation and performance analysis of 3 benchmark models (Standard U-Net, U-Net with Resnet backbone & U-Net with DeepLabV3+ backbone) for Multiclass Semantic Segmentation of Satellite Images.

Primary LanguageJupyter Notebook

Dubai-Satellite-Imagery-Multiclass-Segmentation

Note: Use Jupyter NbViewer to load the code if GitHub doesn't render

Please copy the links of the .ipynb files and paste it in NbViewer if GitHub fails to load it in the browser.

Dataset

Humans in the Loop has published an open access dataset annotated for a joint project with the Mohammed Bin Rashid Space Center in Dubai, the UAE. The dataset consists of aerial imagery of Dubai obtained by MBRSC satellites and annotated with pixel-wise semantic segmentation in 6 classes. The images were segmented by the trainees of the Roia Foundation in Syria.

The images are densely labeled and contain the following 6 classes:

Name R G B Color
Building 60 16 152 #3b1098 #3b1098
Land 132 41 246 #8529f6 #8529f6
Road 110 193 228 #6ec1e4 #6ec1e4
Vegetation 254 221 58 #fedd3a #fedd3a
Water 226 169 41 #e2aa29 #e2aa29
Unlabeled 155 155 155 #9b9b9b #9b9b9b

Benchmark Models

Transfer Learning from InceptionResNetV2 to U-Net CNN

Software & Libraries