Improving flood damage assessments in data-scarce areas by retrieval of building characteristics through UAV image segmentation and machine learning – a case study of the 2019 floods in southern Malawi

Abstract

Abstract. Reliable information on building stock and its vulnerability is important for understanding societal exposure to floods. Unfortunately, developing countries have less access to and availability of this information. Therefore, calculations for flood damage assessments have to use the scarce information available, often aggregated on a national or district level. This study aims to improve current assessments of flood damage by extracting individual building characteristics and estimate damage based on the buildings' vulnerability. We carry out an object-based image analysis (OBIA) of high-resolution (11 cm ground sample distance) unmanned aerial vehicle (UAV) imagery to outline building footprints. We then use a support vector machine learning algorithm to classify the delineated buildings. We combine this information with local depth–damage curves to estimate the economic damage for three villages affected by the 2019 January river floods in the southern Shire Basin in Malawi and compare this to a conventional, pixel-based approach using aggregated land use to denote exposure. The flood extent is obtained from satellite imagery (Sentinel-1) and corresponding water depths determined by combining this with elevation data. The results show that OBIA results in building footprints much closer to OpenStreetMap data, in which the pixel-based approach tends to overestimate. Correspondingly, the estimated total damage from the OBIA is lower (EUR 10 140) compared to the pixel-based approach (EUR 15 782). A sensitivity analysis illustrates that uncertainty in the derived damage curves is larger than in the hazard or exposure data. This research highlights the potential for detailed and local damage assessments using UAV imagery to determine exposure and vulnerability in flood damage and risk assessments in data-poor regions.

Flood Damage Assessment: A Review of Microscale Methodologies for Residential Buildings

“…The multivariate models performed better than traditional damage models. To resolve the challenge of insufficient data in data-scarce regions of Malawi, Wouters et al [41] utilized an object-based image analysis (OBIA) of high-resolution unmanned aerial vehicle (UAV) imagery to extract building characteristics at an individual level and evaluated building damage using local depth-damage curves.…”

“…Although remote sensing is incapable of quantifying damage on its own, it has proven to be efficient when combined with other techniques and tools, such as machine learning [84]. Recently, high-resolution UAV images have been used by the FDA to obtain basic information on buildings [41]. However, it has become common for aerial and satellite images to be manually digitized and labeled to create building footprints, roads, and other utilities [39].…”

“…The data on the buildings are the characteristics of the buildings in terms of their features and geometry. Most building inventory data are sourced from field surveys, local authority databases and buildings records (online and offline), cadastral maps, google street maps, and other such primary databases [10,36,[41][42][43]46,48,[50][51][52][53]93]. Essential to the microscale FDA, building data should be detailed to provide higher and more accurate flood damage predictions [41,54].…”

“…Most building inventory data are sourced from field surveys, local authority databases and buildings records (online and offline), cadastral maps, google street maps, and other such primary databases [10,36,[41][42][43]46,48,[50][51][52][53]93]. Essential to the microscale FDA, building data should be detailed to provide higher and more accurate flood damage predictions [41,54]. The extracted data include building locations, area, the value of the building in currency, flood zone, occupancy type, year built, first-floor elevation, age, and level of maintenance [10,58,97].…”

“…However, various researchers [10,23,[38][39][40][41] have noted that there is a persistent need to bridge the gaps in achieving a reliable and accurate damage model. Such impediments include limited data availability; damage variations within building classes; limited knowledge of the model building, geometries of the building and flood parameters; uncertainties, and validation of models.…”