Discussion Sep 27th

https://arxiv.org/pdf/2102.11958.pdf

Satellite Imagery

SWIR NIR RED 這三個channel比較能凸顯建物 https://developers.planet.com/docs/apis/data/sensors/

Google Earth Engine

Sentinel 2: https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR#bands

SpaceNet

Challenge 1: https://spacenet.ai/spacenet-buildings-dataset-v1/ Band information: https://developers.planet.com/docs/apis/data/sensors/

OpneStreetMap

OpenStreetMap 的好處是公共使用的授權 Google Earth 解析度比較好，但是我們沒辦法合理取得公開上網站的授權，可是小型學術研究應該可以 要看最多違章建築的話，可以直接看彰化鹿港頂番婆

OpenStreetMap API: https://overpass-turbo.eu/

Current plans

• Plan 1: Use the pretrained models from SpaceNet challenges and test them on the satellite imagery in our targeted locations. Potential challenges include: (i) different band information (see above and the band information on Sentinel 2), (ii) generalization from one location to other location
• Plan 2: Use the most recent OpenStreetMap to train a segmentation model from scratch. Potential challenges include: (i) unsure quality of openstreetmap data in Taiwan.