Diverse paradigms of residential development inform water use and drought-related conservation behavior

The first step in our analysis was to gather, process, and integrate multiple data sources using text-matching algorithms, geocoding, and common identifiers. We aggregated and averaged (1) parcel-level housing information from Zillow (Zillow 2017) to define the built environment and (2) Census block-group level demographic information from the U.S. Census and American Community Survey (SimplyAnalytics 2017) to quantify social structure at the block-group level, the smallest spatial scale of Census information. Our final database included 14 features (7 from Zillow and 7 from the U.S. Census), centered and scaled, averaged at the block-group level (n = 46) which were spatially distributed throughout the service area. We calculated average monthly customer-level water use for each block-group from 2008–2017. See the supplemental information for more detailed information on the input data, our cleaning, filtering, and combining procedures, and the final characteristics of the database.

From our final database, we generated community clusters through a sequence of unsupervised learning methods. The built environment and sociodemographic features were highly correlated which prompted us to employ a dimensionality reduction technique. We performed principal component analysis (PCA) to transform the features into a set of linearly uncorrelated variables, thereby compressing our data into a smaller number of variables that explain most of the variation in the feature space and making the clustering simpler and more intuitive. Transforming a set of correlated features into principal components (PCs) to better understand customer profiles has been used across resource sectors, including in water demand analyses both for load shape clustering (Cominola et al 2019) and in regression models (Polebitski and Palmer 2013). We made a scree plot to determine the optimal number of PCs to retain for clustering and identified an elbow at the 5th component. These five PCs together represented 89.6% of the variance in the data (see supplemental information (https://stacks.iop.org/ERL/15/124009/mmedia)).

The PCs were then used as inputs into our clustering algorithm to organize the block-groups into communities. We used the hierarchical clustering on principal components (HCPC) method (Husson et al 2010), which is a combination of partitional and hierarchical clustering. First, a hierarchical clustering procedure is performed using Ward's criterion based on the Euclidean distance. The tree is manually or optimally cut to form an initial set of clusters. Then a partitional clustering method, in this case k-means, is applied to improve the initial groupings obtained from the hierarchical clustering. We applied HCPC to the 46 block-groups and their values for the first five PCs. Based on the intra-cluster inertial plot and values, we determined that five clusters was the optimal solution (see supplemental information). We performed PCA and HCPC using the Factominer package in R (Lê et al 2008). We classified a feature as important to a principal component (PC) based on its contribution: if each variable uniformly contributed to each PC, the expected value would be (1 feature)/(14 total features)*100 = 7%, and we used this 7% cutoff for a feature to be considered important for a PC (Kassambara 2017).

We repeated our principal component and clustering procedure without the Zillow data, using only the 7 Census features to generate community clusters. By creating a 'counterfactual' scenario in which urban environment information is not available, we could compare water use distributions to determine the added value of incorporating built environment features in neighborhood analyses. In this case, we identified an elbow at 3 PCs which were then used for clustering. Based on the intra-cluster inertial plot and values, we determined that five clusters was again the optimal solution (see supplemental information).

We evaluated cluster water use and conservation quantities and trends to evaluate how different residential neighborhood typologies are linked to water demand. Monthly block-group water use and summer monthly block-group water use distributions within each cluster were not normal (Shapiro–Wilks test, all 10 tests p < 0.05), thus we used the nonparametric Kruskal–Wallis test to statistically compare water use distributions between community clusters. Tests for (1) all months and (2) summer months were significant (p < 0.05) indicating that water use distributions across groups were not statistically similar. To tease out individual group differences, we used Dunn's post-hoc test with the Benjamini–Hochberg method of adjusting p-values to minimize false discovery rates.

We examined block-group water conservation and rebound between 2014 and 2017, a drought stress-test, which provides insight into how communities react to an external system shock, to capture changes in water use related to extreme drought, political actions, and local policies within Redwood City (table 1). In January 2014, the California Governor declared a drought state of emergency and called for voluntary water reductions across the state (California State Water Resources Control Board 2014). As the drought progressed, the state issued a resolution calling for conserving potable outdoor water use (California State Water Resources Control Board 2014) which coincided with Redwood City implementing outdoor water use restrictions for single-family residential customers. Customers were limited to 2–day a week outdoor watering where residential addresses ending with odd numbers were allowed to water on Monday and Thursday and with even numbers water on Tuesday and Friday. Following a statewide drought declaration in April 2015, June 2015 brought state-level mandatory water use restrictions where Redwood City was mandated to achieve 8% overall savings (California State Water Resources Control Board 2018); however, the city did not change their single-family residential irrigation policy, only the level of enforcement. The outdoor water use restriction ended in June 2016 when the mandatory conservation regulations ended and a state-wide call for 'self-certified conservation goals' at which time Redwood City set their goal to 0% (California State Water Resources Control Board 2018). The drought was declared over in April 2017.

We calculated monthly block-group absolute and relative (percentage) water conservation rates compared to the same month in 2013 in order to follow California's statewide drought mandate requirements (California State Water Resources Control Board 2015). The absolute and relative block-group level conservation rate distributions had similar features as the water use distributions (were not normal, Shapiro–Wilks test, all 10 tests p < 0.05; were not statistically similar, Kruskal–Wallis test, both tests, p < 0.05) so we deployed the same statistical testing procedures for the group comparisons. We compared conservation rate distributions between clusters for each of the five distinct policy periods.

To investigate the importance of the built environment in grouping customers for water use analyses, we also created clusters trained only using Census data. We compared cluster cohesion, or within cluster sum of squares (WCSS), between the two solutions for four different water-related variables: water use, water use with seasonality removed, absolute water conservation, and percent water conservation, which were not inputs into the clustering algorithms:

$$\begin{equation}\mathop \sum \limits_i^{} \mathop \sum \limits_{x \in {C_j}}^{} {\left( {x - {{\underset{-}{x} }_i}} \right)^2}\end{equation} \tag{ 1 }$$

where the Euclidean distance is calculated between each point $x$ and the mean ${\text{ }}\underset{\scriptscriptstyle{-}}{x}$ in cluster i and then squared and summed across all clusters. A smaller WCSS indicates tighter cohesion and thus a clustering result that is more useful for water use and conservation analyses.

Five residential community clusters emerged from our analyses. One important outcome is that although no geospatial inputs i.e. location coordinates were introduced into our model, the clusters were separated into spatially distinct groups (see supplemental information). This unprompted spatial segregation provides evidence for both geographical sorting of people into similar networks and community-scale urban planning while showing the effectiveness of clustering by built environment and social features. This kind of spatial grouping also paves the way for neighborhood-level distributed water system integration, for example decentralized water recycling infrastructure and geospatially tailored customer outreach (Gurung et al 2016).

The features of each cluster can be examined by spider plots of feature value z-scores (figure 1) and average feature values (table 2). Cluster A is comprised of 11 block-groups and formed by block-groups/neighborhoods with generally low-income, high-density families with many people per household. Cluster B is made up of 12 block-groups and is the most similar to Cluster A in that neighborhoods in this community represent lower income customers. However, unlike Cluster A, block-groups are linked to nonfamily renters with bigger household sizes. Cluster C is made up of 11 block-groups and represents the average of Redwood City across the social and built environment.

Zoom In Zoom Out Reset image size

An especially interesting result is the division of high-income customers into two distinct groups. Cluster D and Cluster E both contain block-groups with high positive values in PC1, representing affluent, highly educated neighborhoods with bigger houses. Cluster D, containing 4 block-groups, however, represents an area with younger residents living in newer houses on smaller lots with more people per household. These characteristics are in contrast to the community of Cluster E, containing 8 block-groups, which is dominated by older houses on large lots with less people per household.

Aerial imagery provides evidence for these neighborhood typologies (figure 1). For example, these pictures show dense development for Clusters A, B, and D, although the development in Cluster D shows a planned community with shared instead of individual lawns. Imagery from Clusters C and E both show spaced out single-family residential properties. The lots in Cluster E are substantially larger than those in Cluster C, however, and most include lawns.

What do these neighborhood clusters mean for water use patterns and trends? A cycle plot shows average monthly water use within each cluster for the entire 2008–2017 time period (figure 2(A)) with average water use and summer water use values presented in table 3. To better understand the seasonality shown in the cycle plot, we calculated peaking factors (Polebitski and Palmer 2010) for each block-group as the ratio of average monthly water use in the summer months divided by average monthly water use in the winter months (figure 2(B), table 3). All block-groups have average peaking factors greater than 1, indicating at least some outdoor water use within all single-family neighborhoods throughout the city. Additionally, block-group peaking factors within each cluster were not widely distributed, demonstrating the link between neighborhood typologies and outdoor water use.

Zoom In Zoom Out Reset image size

We present boxplots within violin plots, which are boxplots with rotated kernel density shapes, to visualize average monthly and average monthly summer water use by customers in block groups within each cluster (figures 2(C) and (D)). Statistically, eight of the ten water use distribution comparisons were significantly different (supplemental information) while all ten summer water distribution comparisons were significantly different (supplemental information), indicating overall different water use among the clusters. The two comparisons that were not statistically different were between Clusters A and C and between Clusters B and D. The cycle plot sheds light on these connections. Clusters B and D do in fact show similar seasonal use, in contrast to Clusters A and C, where Cluster A, with older houses and more residents per household, has higher winter use and Cluster C, with more homeowners (compared to renters) and bigger lot sizes, has higher summer use.

Block-groups in Cluster A had medium monthly and summer monthly water use compared to the other clusters and the lowest average peaking factor of 1.15 (a value shared by Cluster D). Cluster A is also the lowest income community, with high density housing, the most people per household, and mostly renters. The other low-income, majority renter community, Cluster B, had a similar mean summer water use but lower winter water use and therefore a slightly higher peaking factor of 1.21. This divergence could be due to a higher percentage of families, more people per household, and more multi-unit (although still considered single-family residential) buildings in neighborhoods that define Cluster A compared to Cluster B. Cluster C, the average of the service area across most features, exhibits the second highest monthly water use, summer monthly water use, and peaking factor of 1.28, indicating room for increased outdoor conservation and efficiency. Interestingly, Clusters C and D have similarly low winter water use to that of Cluster B, pointing towards indoor efficiency within these higher-income groups.

Cluster D had the lowest summer water use and also the lowest peaking factor of 1.15. If we only examined block-group demographics, this would be a surprising result as neighborhoods in this community are characterized by high-income, highly-educated homeowners, with larger household sizes, and these features are typically associated with higher water use (Harlan et al 2009, Polebitski and Palmer 2010, House-Peters et al 2010, Mini et al 2014, Breyer et al 2018). However, by incorporating features of the built environment, extracted from Zillow, we can understand that low water use is linked to new urban form, built in the 1990s compared to the 1940s–1960s like the rest of the city. Additionally, through aerial imagery (figure 1), we found that within the high-income neighborhoods of Cluster D, many areas have substituted individual lawns for community or homeowner's association lawns. While water was still being used for outdoor irrigation around these houses, water use for shared lawns is not included in this study. Water use decisions made at the individual homeowner (or renter) level are fundamentally different than those made by homeowner associations and landscape professionals. In addition, many of these shared lawns are being irrigated with recycled water, making them one avenue for more efficient green spaces in urban settings (Quesnel and Ajami 2019).

Finally, customers in Cluster E had the highest average water use and highest average summer water use, including the highest mean peaking factor of 1.35, compared to the other cluster communities, which can at least be partially explained by the large lots and large houses. Notably, Cluster E contains one outlier block-group which has substantially higher income, water use, and lot sizes compared to the rest of the service area and the other block-groups in the cluster. The other six block-groups within Cluster E do, however, have the next highest water use profiles, thus still representing a cohesive high water-use group.

Customers were able to achieve high conservation rates, especially as the drought progressed (table 4, figure 3). During the voluntary conservation period in the first half of 2014, customers in each cluster conserved at similar relative rates, with average cluster monthly block-group savings between 6%–12% compared to 2013 across the service area. As the drought progressed and statewide outdoor water use restrictions were coupled with local watering policies, average cluster monthly block-group savings increased to 14%–23% compared to 2013. Throughout these first two policy periods, the three lower water use clusters A, B, and D had similar absolute savings while the two higher water use clusters C and E conserved in parallel and had the highest relative savings.

Zoom In Zoom Out Reset image size

In spring 2015, mandatory water use restrictions were implemented across the state which led to peak drought awareness during the 2014–2017 period (Bolorinos et al 2020). This mandate, coupled with the continuation of local watering restrictions, resulted in the policy period with the highest relative and absolute conservation rates within each cluster during the drought, with average cluster monthly block-group savings of 23%–37% compared to 2013.

When the statewide and local mandatory water restrictions were lifted, the 'self-certified' goal policy period went into effect. Customers maintained high average cluster monthly block-group conservation rates of 20%–24% compared to 2013 despite the absence of restrictions. However, in the spring of 2017 when the drought was declared over, conservation rates lessened to 11%–25% compared to 2013.

From the beginning of the drought until the self-certified goals were lifted, Clusters A and D had similar relative and absolute conservation rates, which were the lowest in the service area. These two groups have dramatically different demographic and housing features: Cluster A is comprised of low-income renters in small, older houses compared to Cluster D which contains block-groups defined by high-income homeowners in large, newer houses. However, both have small-lot sizes, the lowest peaking factors and therefore low household outdoor water use. One notable difference occurs when the drought ends. Clusters A and B, with similar demographic and built environment profiles, have similar water use rebound responses while Cluster D exhibits the highest rebound (lowest conservation rates) in the service area, indicating different community responses to restrictions being lifted.

During every policy period, customers in the two communities with the highest overall water use, summer water use, and highest peaking factors, Clusters C and E, had the largest absolute savings, signaling the high potential for outdoor water conservation. During later drought periods, customers in Cluster E were able to achieve higher absolute savings, however, likely linked to their higher summer and overall water use. As urban landscapes can often recover after drought periods, even with decreased irrigation (Quesnel et al 2019), there is a strong case for limiting outdoor water use during water supply shortages.

Researchers have consistently reported that affluent communities are responsible for the greatest savings during drought because they generally have the largest amount of outdoor space, and thus outdoor water use (Kenney et al 2008, Mini et al 2015, Breyer et al 2018). Yet in this study, we found that Clusters D and E were both defined by high-income, highly educated customers but had dramatically different water use behavior during and after drought due to their built environment factors. This outcome demonstrates the importance of taking into account differences in urban form when assessing and forecasting water use, including when developing drought plans.

We found that across all four cluster comparisons, the clusters created using both Zillow and Census features had lower WCSS (table 5). The clusters that included Zillow showed a 15% improvement in overall water use cohesion and a 30% improvement in de-seasoned water use cohesion, pointing to the importance of including urban form not only when assessing total water quantities but also for examining water use trends and behavior over time. Both water conservation WCSS metrics also improved for the cluster solution that included Zillow, although to a lesser extent at 11% and 3% performance increases.

The cluster solution that takes into account both built environment and sociodemographic features splits the high-income customers into two distinct clusters (figure 1) leading to a logical grouping of their water use patterns (figure 2) and conservation and rebound behavior (figure 3). Comparing block-group water use and income relationships for both cluster solutions (figure 4) further highlights the important decoupling between income and water use, which comes through in the cluster solution only when built environment features are included.

Zoom In Zoom Out Reset image size

The two cluster solutions diverge the most specifically when examining the higher-income block-groups. In the first cluster solution that includes features of the built environment (figure 4(A)), the higher-income customers are stratified into three groups that also correspond to three different levels of water use, despite no water use characteristics being used to generate the clusters themselves. This is especially important for the highest-income customers, where the block-groups are separated into three groups—Cluster E with the highest water use, Cluster C with high to medium water use, and Cluster D with lower water use. However, in the cluster solution with sociodemographic features only (figure 4(B)), the high-income customers are all grouped into one neighborhood typology, Cluster Z, but water use spans the range of the entire city. The traditional method of classifying customers and neighborhood based only on affluence will not be sufficient for understanding water use trends and tailoring various demand management strategies as new types of urban development are integrated into cities. This distinction has important consequences for demand forecasting and supply planning. If water agencies over-predict water use from high-income customers, they will be unable to optimally invest in diverse water supplies and will mis-allocate resources throughout the community.