A Generative AI Model Of Urban Spaces In The Face Of Climate Change

Urban land use structures have impact on local climate conditions.To shed light on the mechanisms of local climate w.r.t.urban land use,we present a data-driven,deep learning approach based on airborne LiDAR data statistics and the Landsat 8 satellite’s surface temperature product.Our study proposes a deep neural network architecture and data workflow to model the following question: How to vary a geospatial scene’s ambient temperature by modifying without bias its urban land use structures represented by LiDAR statistics?We model the phrase modify without bias by a constraint that allows all LiDAR statistics features to contribute to modelled temperature variations at same order of magnitude.In contrast to regular deep learning neural network optimization,we consider fixed model parameters,but variation of model input data,i.e.LiDAR statistics.In more detail,we utilize the 2017 LiDAR survey of New York City to generate raster layers based on LiDAR laser pulse characteristics.For example,we employ elevation information(laser light’s time-of-flight measurements),the laser pulse return count,and the reflected laser light’s intensity in order to generate gridded,spatial statistics.As demonstrated in the literature,such statistics bear signature of human infrastructure such as buildings,vegetation,and traffic networks.For a model t(s)correlating LiDAR statistics s with ambient surface temperature t≥ 0(in Kelvin),a lower surface temperature inferred by dt(s)/ds and Δt<0 corresponds to gradient descent1,while(incrementally)increasing Δt>0 resembles gradient ascent of model t(s).Instead of naively backpropagating a temperature shift Δt through model t(s),we employ a variational autoencoder E(?)D to generate encoded,compressed feature vectors c from the LiDAR statistics s minimizing |s-D(c)| where c=E(s).This way,variations in c produce smooth variations in the regenerated LiDAR statistics s’=D(c).Linking a mean Landsat 8 temperature value t to the geolocation covered by s(and associated with c),we train a regression model R correlating c and t,R(c)=t.Subsequently,we add a shift Δt to the temperature t of a scene represented by c=E(s),and backpropagate to c’=c+Δc.We decode the modified feature vector c’ into modified LiDAR statistics s’=D(c+Δc).Finally,we apply to s’ the concept of Auto Geo Label in order to estimate the change in the fraction of vegetation of an urban scene associated with t→t+Δt.The novelty of this thesis comprises of the introduction and evaluation of a deep neural network architecture that correlates vegetation and ambient temperatures from remote sensing modalities.The concept helps to approximate the climate resilience of urban areas.By analyzing numerous vegetation vs.temperature change pairs,we develop a statistical evaluation procedure to perform correlation analysis.The approach generates a qualitative answer to the question posed above:When statistically averaged over New York City with focus on the Queens borough,an increase in vegetation correlates with a decrease in ambient surface temperature with likelihood of 95%.Our contribution likes to inspire the development of urban heat island mitigation strategies in the face of climate change.