A data-driven approach to rapidly estimate recovery potential to go beyond building damage after disasters

Natural hazards disproportionately impact vulnerable populations, exacerbating inequality for years. Recovery policies often fail to address this, prioritizing aid based on pre-disaster assets (loss-based assistance) rather than needs. While needs-based approaches exist, their implementation is hindered by a lack of timely information identifying communities likely to experience impeded recovery. Current post-disaster data collection overwhelmingly focuses on quantifying building damage, a myopic measure of long-term recovery needs. This study proposes focusing on 'non-recovery,' specifically those who fall behind in recovery over time, to identify communities with disproportionate needs. A data-driven approach is developed using readily available data (census, remote-sensing, modeled data) to estimate non-recovery, demonstrated using housing reconstruction data from the 2015 Nepal earthquake. This approach allows for a rapid assessment of recovery potential, identifying regional differences and vulnerabilities that may not be apparent from damage data alone, ultimately informing more equitable and sustainable recovery strategies.

The paper references several studies highlighting disaster-exacerbated inequality. Examples include the disproportionate impact of the 2004 Indian Ocean tsunami on women and inadequate access to assistance programs for marginalized communities after the 1994 Northridge earthquake. Existing literature emphasizes the importance of housing recovery policies in prioritizing vulnerable populations. However, current approaches are often loss-based, failing to adequately address the needs of those lacking pre-disaster assets. The authors point to the limitations of the Post-Disaster Needs Assessment (PDNA) process, which often focuses on economically quantifiable damages rather than social needs. They also critique the overreliance on remote sensing and digital crowdsourcing data for quantifying building damage, neglecting the broader social and environmental factors influencing long-term recovery. Previous research on Nepal's earthquake recovery highlights the role of various factors in impeded reconstruction, including hazard exposure, accessibility, poverty, and reconstruction complexity. However, these studies often rely on survey data that is not readily available in the immediate aftermath of a disaster. The authors contrast their data-driven approach to existing index-based methods for estimating vulnerability or resilience, highlighting the limitations of these index-based methods, especially in post-disaster contexts in developing countries.

The study uses a data-driven approach to estimate non-recovery after the 2015 Nepal earthquake, focusing on housing reconstruction. The approach involves relating surveyed non-recovery outcomes (in this case, whether households fully completed reconstruction within four and a half years) to predictor variables from readily available data sources. The study area comprises eleven districts outside Kathmandu Valley, selected due to the focus of Nepal's Earthquake Housing Reconstruction Program. Reconstruction progress is measured as a binary variable (completed reconstruction or not). The initial set of 32 predictor variables considered capture sociodemographic, economic, environmental, and geographic factors likely to influence recovery. These variables were selected through a combination of interviews with stakeholders, literature review of the impacts of and recovery from the Nepal earthquake and broader theories on sustainable development, vulnerability and resilience. Data included earthquake shaking intensity, tree cover, population density, remoteness, landslide hazard, access to tap water, topographic slope and food poverty prevalence. Variable selection was done using an automatic technique to remove less predictive variables, resulting in eight variables for the final model. A random forest model was used to predict the probability of non-reconstruction, chosen for its ability to handle non-linear relationships and interactions between variables. The model was trained and validated using stratified random sampling, with 84% of the data used for training and 16% for testing. The performance of the random forest model was compared to logistic regression, with the random forest demonstrating superior performance. Partial dependence plots were used to visualize the relationships between predictors and the probability of non-reconstruction. Finally, the model was applied spatially to generate a map predicting the probability of non-reconstruction across the study area.

The analysis identifies eight key predictors of non-reconstruction in Nepal, categorized into hazard exposure, rural accessibility and poverty, and reconstruction complexity. Hazard exposure includes earthquake shaking intensity and rainfall-triggered landslide hazard, both significantly associated with increased non-reconstruction probability. Rural accessibility and poverty are represented by remoteness, tree cover, and food poverty prevalence; remote households and those in areas with high food poverty are significantly less likely to complete reconstruction. Notably, higher tree cover above a certain threshold showed a positive association with reconstruction. Reconstruction complexity involves population density, access to tap water, and topographic slope; higher population density and steeper slopes are associated with higher non-reconstruction probability, while a higher prevalence of tap water surprisingly correlates with slower reconstruction, potentially due to more complex logistics in better-connected villages. Spatial analysis reveals that areas with high non-reconstruction probability are not necessarily those with the highest damage; non-reconstruction shows a unique spatial pattern dictated by the interaction of social, geographic, and environmental factors. The model demonstrates good predictive performance (AUC of 0.725 on the test set).

The findings demonstrate the value of moving beyond damage assessments to estimate non-recovery in the aftermath of disasters. The data-driven approach presented allows for a rapid identification of factors hindering long-term recovery and highlights the importance of considering pre-existing vulnerabilities. The results reveal a complex interplay between physical hazards, socioeconomic conditions, and the logistical challenges of reconstruction. The spatial distribution of non-reconstruction differs from that of building damage, emphasizing the limitations of using damage data alone for recovery planning. The study shows that areas with high non-reconstruction probability may be overlooked if only building damage is considered. The model’s results highlight the importance of considering ongoing risks like landslides and food insecurity in recovery planning. The model provides a more nuanced understanding of recovery needs compared to index-based approaches, offering a direct measure of a recovery outcome and capturing the complex interplay between factors influencing it. The model's results could improve aid allocation and support policies that target both highly damaged areas and those with significant pre-existing vulnerabilities.

This study introduces a data-driven approach for estimating non-recovery after disasters, focusing on housing reconstruction in Nepal. The model identifies key predictors across hazard exposure, accessibility/poverty, and reconstruction complexity, highlighting the importance of a holistic view beyond building damage. The findings emphasize the need for incorporating pre-existing vulnerabilities and ongoing risks into recovery planning and aid distribution. The approach offers a more actionable metric than traditional vulnerability indices, providing a rapid assessment tool to improve the equity and sustainability of disaster recovery efforts. Future research could explore different recovery outcomes, expand the range of predictor variables (including granular social data), and apply this approach to other disaster contexts.

The model's accuracy might be affected by the interaction between reconstruction and external aid beyond the government program. The model's generalizability to future earthquakes in Nepal or other regions needs further investigation. The resolution of predictor data limits the model's ability to capture fine-grained household-level variations in non-recovery. The model does not explicitly incorporate some Nepali-specific vulnerability factors (gender, caste) due to data limitations, though these factors are recognized as potentially significant.