Satellite Imagery

Building a Deep Learning model using U-Net architecture for satellite images.

Demo

Satellite Segmentaion Prediction - a Hugging Face Space by immortalPi https://huggingface.co/spaces/immortalPi/Satellite-Segmentaion-Prediction

Documentation

• satellite dataset - dubai
• Convert the image sizes 
• masked segmentation images
• data preparation for deep learning
	○ Patchify 
• U-net model
	○ Encoding and decoder 
• Custom Loss function and metrics 
• Segmentation model
• Training deep learning model
• Model history and prediction
• Activation outputs and gradients for each tensorflow layer 
• Netron - visualize the model
• ML training diagnostics (Local) - https://wandb.ai/site
	○ https://wandb.ai/26-amruth-immortalpi/satellite-segmentation-dubai-images 
• Apply remote debugging with weights and biases 
• Activation output for each layer 
	○ Keras activation + gradients  Keract
	○ Layer wise visualization 
• Model deployment at hugging face

API Reference

Weights and biases

  https://wandb.ai/26-amruth-immortalpi/satellite-segmentation-dubai-images

Parameter	Type	Description
api_key	string	f6696212866a9b097f6468d5cae2901364cf****

hugging Face

  https://huggingface.co/spaces/immortalPi/Satellite-Segmentaion-Prediction

Parameter	Type	Description
upload images	image	upload custom images to check

Screenshots

objective

model

https://netron.app/ (load the model for viz)

weights and biases

hugging face

Optimizations

needs Optimizations

custom loss function

cross entropy loss extension