developmentseed/looking-glass repository overview

⁠Overview

Looking Glass is a building segmentation model for identifying buildings from zoom⁠ 18 satellite imagery. The model is based off of Google's DeepLabV3 and was trained on all 30cm data from the SpaceNet building training set⁠. The model is a "generalist" in that it wasn't tailored to any specific region -- the optimization goal was to perform well across a variety of geospatial regions.

⁠Inputs and Outputs

The docker image takes in a json-like dict (via a POST request) with base64 encoded images. Images should have been read from a standard format (like png or jpg) where the valid pixel range is 0-255. When transforming the 0-255 range to 0-1, Looking Glass will perform some histogram equalization to set the minimum pixel value to 0 and the 98th percentile pixel value to 1. See serving_input_receiver_fn⁠ in the Looking Glass repo for more details on how the model preprocesses images.

Looking Glass will return a json response with the float values corresponding to the signed-distance transform. Predicted pixels outside a building have a value < 0, pixels on a building border have value 0, and pixels inside a building have values > 0. It's suggested to use a threshold value of >=0 to binarize the building prediction mask.

Note: be sure your images are projected into an equal area projection if you want to sum up pixels and get the predicted square meters of building area.

⁠Example

To see a full example, see the jupyter notebook walkthrough⁠ in the looking-glass repository.

⁠Raw model files

The raw Keras weights and parameters used to build this image are available and stored by the Humanitarian OpenStreetMap team:

• weights: https://s3.amazonaws.com/ml-enabler-keras/0524_215014_weights.h5⁠
• params: https://s3.amazonaws.com/ml-enabler-keras/0524_215014_params.yaml⁠