Home

Awesome

UNET-Image-Segmentation-Satellite-Pictures

Predict housing rooftops on satellite pictures

Description:

Unet to predict roof tops on satellite pictures. Training the model was done with 10000 satellite pictures from the Crowed AI Mapping Challenge. Mask created with MS Coco annotations. Project includes automatic Mask creation for Train and validation set, Unet model with optimizer, data processing including a Generator function and a function for creating plots. Pretrained weights are also included from which the following plots were created.

Example1 Example2

(see also Example_Plots)

Prerequisites:

Done with anaconda environment, see environment.yml for details.

Data:

Model was trained and evaluated with data from the crowed ai mapping challenge. Moderate data augmentation was used: horizontal and vertical flips. Data Augmentation was added the in the generator and randomly executed, see dataprocessing.

Data_Workflow

Model:

Unet model (Decoder - Encoder Network) with Adamax optimizer, loss is calculated by binary crossentropy.

Example2

(Unet picture from the original paper. Unet in the present projects differs)

U-Net consists out of 2 seperate networks, one decoder and one encoder. Example3

The backbone of both are convolutional layers, but the technics differ. While the decoder extracts features with its kernels... Example3

...the Encoder generalize from these features to create the feature mask. Example3

Get Started:

Good start includes: