A Probabilistic Classification Framework for Urban Structural Damage Estimation Using Satellite Images

Recent research endeavors in civil engineering have attempted to apply remote sensing technology to urban damage estimation following large-scale urban disasters. Different from a general change detection problem, wherein categorical or mostly binary change states (`changed' or `unchanged') are associated with pixels, urban structural damage is usually described on a per-object basis using qualitative damage states, such as `fully collapsed', `partially collapsed', or `intact' for individual structures or sub-areas. In addition to this difference, another limitation in past attempts of image-based urban damage estimation is the use of a deterministic approach to estimate damage states, such that resulting damage statistics have no confidence levels. To address these limitations, probabilistic learning theory is applied in this paper to estimate structural damage on a per-object basis. The result is individually categorized structures or sub-areas in terms of different damage states, which can quickly be assessed via color-rendered damage maps. Such a result, particularly where probabilities are associated with the estimated damage states, is useful for post-event reconnaissance and recovery.