Automated Generation of Building Footprints in Satellite Imagery using Artificial Intelligence

Oregon Convention Center 

A deep learning algorithm is implemented to identify building footprints in satellite or arial imagery. A Mask Region-Based Convolutional Neural Network (Mask R-CNN) was trained and tested for detection of residential buildings. Annotated dataset for training was generated by sketching bounding boxes across the buildings. Resnet50 was used as the backbone for transfer learning in the model for detection of building footprints. The dataset for re-training and fine-tuning of the transfer network was the 2024 Bay County building footprints for annotation of the 2022 NAIP images taken over Panama City, Florida. The training data was created by making 2851 image chips of size 256 x 256 from the original image using the Bay County footprints as the labels. The learning rate for the model was 1.0000e-04 with an Average Precision Score of 0.72. It was then tested on images in Mexico Beach and other areas of Panama City for a final test and the results were promising. The unregularized footprints show good agreement with the current image used for inference. Looking at the confidence in the predictions, the lowest level recorded in the attribute table was 0.93.

GIS

Artificial Intelligence

Satellite Imagery

Building Footprint

Object Detection

Georeferencing 

Section on Statistics in Imaging