Damage classification and segmentation

Task description

Via computer vision and exclusive datasets made by drones, we need the neural network pipeline the best quality methods for damage classification and segmentation

Guideline

We classify 4 types of buildings on the map using the next guideline Destruction Guideline | Ukraine War: 

1. 1. Destroyed: assigned to structures that are total or largely collapsed (>50%). This category shall be assigned also when only a portion of the building has collapsed to the ground floor. In these cases, the original building structure is no longer distinguishable.
   2. Major Damage: it shall be used when post satellite imagery is available and includes: Major visible damages, which shall be assigned to structures with part of the roof collapsed and serious failure of walls; Minor visible damage level, i.e. buildings with a largely intact roof characterised by presence of partial damage (collapse of chimneys or roof tiles detach) or surrounded by large debris/rubble or sand deposit
   3. Minor Damage: it shall be used for buildings whose interpretation is uncertain, due to lower image quality (e.g. shadow or degraded resolution due to high off-nadir angle) or to the presence of possible damage proxies like small traces of debris/rubble or sand deposits around the building. This class attribution can be given by inferring the state of the building from surrounding features.
   4. No Visible Damage: structures with no visible from photo damage, it shall be assigned to the structures that appear to have complete structural integrity, i.e. when the walls remain standing and the roof is virtually undamaged. It is important to remark that this class don’t exclude the presence of structural damages, i.e. the building may anyway have suffered damages that can’t be assessed from vertical satellite imagery regardless of its spatial resolution.

Datasets examples

Full Datasets will be available to the validated users, please signup in our Marketplace to be able collaborate with our clients