DeepC4: Deep Conditional Census-Constrained Clustering for Large-scale Multitask Spatial Disaggregation of Urban Morphology

TL;DR

Deep Conditional Census-Constrained Clustering (DeepC4), a novel deep learning-based spatial disaggregation approach that incorporates local census statistics as cluster-level constraints while considering multiple conditional label relationships in a joint multitask learning of the patterns of satellite imagery is presented.

Abstract

To understand our global progress for sustainable development and disaster risk reduction in many developing economies, two recent major initiatives - the Uniform African Exposure Dataset of the Global Earthquake Model (GEM) Foundation and the Modelling Exposure through Earth Observation Routines (METEOR) Project - implemented classical spatial disaggregation techniques to generate large-scale mapping of urban morphology using the information from various satellite imagery and its derivatives, geospatial datasets of the built environment, and subnational census statistics. However, the local discrepancy with well-validated census statistics and the propagated model uncertainties remain a challenge in such coarse-to-fine-grained mapping problems, specifically constrained by weak and conditional label supervision. Therefore, we present Deep Conditional Census-Constrained Clustering (DeepC4), a novel deep learning-based spatial disaggregation approach that incorporates local census statistics as cluster-level constraints while considering multiple conditional label relationships in a joint multitask learning of the patterns of satellite imagery. To demonstrate, compared to GEM and METEOR, we enhanced the quality of Rwandan maps of urban morphology, specifically building exposure and physical vulnerability, at the third-level administrative unit from the 2022 census. As the world approaches the conclusion of many global frameworks in 2030, our work offers a new deep learning-based mapping technique that explicitly encodes well-validated census and experts' belief systems to achieve an explainable and interpretable auditing of existing coarse-grained derived information at large scales.

Figures (7)

• Figure 1: Geographical extent of Rwanda gadm. The boundaries of sectors and provinces are respectively displayed with thinner and thicker lines. Every distinct color represents a district.
• Figure 2: DeepC4 Implementation. From left to right, we began with the encoding of census conditional constraints, preparations of EO features, and selection of disaggregated building locations from various imperfect sources. We then trained an autoencoder to obtain a set of reduced-dimension latent representations that were used for constrained clustering algorithms. Jointly trained using the reconstruction and prediction losses, we evaluated the performance of the autoencoder using the available building-level groundtruth.
• Figure 3: Comparison of the spatial disaggregation between DeepC4 and METEOR at 500-meter pixel. Calculated percentage differences (a) visualized on a map and (c) expressed as a histogram chart showing the proportion of pixels per province. Positive percentage differences indicate that outputs of DeepC4 are larger than those of METEOR and negative otherwise. Categorical difference as (b) a map and (d) a histogram chart showing the proportion of pixels for each descriptive category. The color encoding of map and bar chart is shared between (a) and (c), and between (b) and (d).
• Figure 4: Learning curves. (a) reconstruction loss (i.e., how effective the autoencoder is in reducing the high-dimensionality into a meaningful reduced-dimension representations); (b) prediction loss (i.e., how accurate the resulting reduced-dimension representations are with respect to available building groundtruth info); (c-e) absolute $p_{valid}$ for each prediction task (roof, wall, and height); and (f-h) similar plots but accounting for the imbalanced number of target classes.
• Figure 5: Usefulness of partially available building-level groundtruth. (a-d) a set of prediction maps as a result of DeepC4; (e-g) corresponding $p_{valid}$ maps; (h) reference background map with sector labels; (i-k) aggregated sector-level values of $p_{valid}$ maps; and (l) reference location.