Addressing Challenges of Deep Learning for Post-Disaster Damage Classification

The effectiveness of natural disaster response functions has increasingly relied on information extracted from remotely-sensed imagery – particularly for the time-critical and often large-scale task of assessing post-disaster damage to the built environment. Despite a recent push to operationalize deep learning (DL) technologies to assist this task, several overlooked and unaddressed challenges limit DL’s applicability. First, very little DL research in the remote sensing domain utilizes the geographic variation in objects of interest or ancillary spatial data layers as features which could help the algorithms learn. Second, there is a scarcity of labelled training data depicting the effects of common natural hazards on all operationally-meaningful types of infrastructure. Third, the label noise and intra-class variation present in existing datasets – due to a mostly non-expert and subjective labeling process – are generally not taken into account. This presentation will discuss how these challenges may be overcome through geospatial feature engineering in the DL framework, generative modelling for dataset augmentation, and fuzzy representations of data labels.