Overcoming built data-scarcity in developing cities: Hidden Markov Methods to construct reliable building footprint data across urban climate risk zones

Abstract Prospective climate risk assessments for climate change adaptation and emergency management rely on reliable, accurate data about the built environment. Yet, urban areas in developing countries are growing rapidly, so data sources and methods that measure urban growth in a timely manner are critical. However, current methods that leverage satellite data and machine learning to produce building footprint datasets are prone to biases correlated with urban risk due to limited training data across different continents and types of urban areas, as well as challenges in interpreting satellite imagery across different urban forms. In this paper, we aim to improve the reliability of building footprint data across urban forms through the integration of limited local data using Hidden Markov Models. We present three key contributions: 1) An urban climate risk assessment framework to evaluate datasets derived from deep machine learning models and satellite imagery across urban forms; 2) A method for processing probabilistic outputs of aggregate building footprint data to account for uncertainty among risk classes; 3) A Hidden Markov Model method to calibrate CNN outputs in post-processing with small local datasets to overcome biases critical to climate risk assessments and downstream management decisions. In a case study of Kigali, Rwanda, we show that Hidden Markov Models calibrated on data from similar Local Climate Zones can improve the MSE of built area percent at a block scale from the current building footprint models at 6.8% down to 2.4%. Furthermore, these models reduce standard deviation in performance of estimation of percent built area across Local Climate Zones from 6.6% to 2.6%, reducing the variability in the reliability of built area estimates in high-risk Local Climate Zones.