Skip to content

Home

Jump to bottom
Annie edited this page Aug 17, 2018 · 41 revisions

Disclaimer

This repository is modified from the code for the the DIUx xView Detection Challenge. The paper is available here.

This repository is created for Automatic Damage Annotation on Post-Hurricane Satellite Imagery, one of three projects from the 2018 Data Science for Social Good summer fellowship at the University of Washington eScience Institute.

Introduction

Two object detection algorithms, Single Shot Multibox Detector and Faster R-CNN were applied to satellite imagery for hurricane Harvey provided by DigitalGlobe Open Data Program and crowd-sourced damaged buildings labels provided by Tomnod. Our team built dataset for damaged building object detection by placing bounding boxes for damaged buildings whose locations are labelled by Tomnod. For more information about dataset creation, please visit our website.

We usedtensorflow object detection API to run SSD and Faster R-CNN. We used a baseline model provided by xView Challenge as pre-trained model for SSD. This repository contains code performs data processing training, specifically, 1) image chipping, 2) bounding box visualization, 3) train-test split, 4) data augmentation, 5) convert imagery and labels into TFrecord, 6) inference, 7) scoring, 8) cloud removal, 9) black region removal, etc. See here. We fed the data into SSD and Faster R-CNN and predicted the bounding boxes for damaged and non-damaged buildings.

Data Processing

The dataset contains 875 satellite imagery of the size 2048 x 2048 in geotiff format. Each of the imagery contains at least one bounding box. The dataset also includes a geojson file containing bounding box location and its class (damaged building or non-damaged building). The dataset contains many invalid bounding boxes that do not cover any building, for example, bounding boxes over the cloud or in the field. Automatic cloud removal method was applied and followed by manual cleaning. Data was then split into training, validation, and test set.

Imagery were converted to tiff files and chipped into smaller non-overlapping images for training. Different sizes (500 x 500, 300 x 300, and 200 x 200) were experimented and 200 x 200 was chosen because it resulted in better performance. This may because 200 x 200 chips contains less context and buildings appear larger in resultant images than other two sizes. Bounding boxes cross multiple chips were truncated at the edge, among them, those with small width ( < 30 pixels) were discarded. Chips with black region area larger than 10% of its coverage were discarded. Then the chips and its associated labels went though optional processing (i.e., data augmentation) and were converted to TFrecord ready for training and testing.

Model Implementation

Reference:

Clone this wiki locally