Resilience Insights

We report the frequency of severe drought in months per decade because that is a threshold at which the impacts become substantial. In the scPDSI, negative values indicate drought, and a value of -3 is considered severe drought. We use the CRU scPDSI historical data from 1900-2020 to represent current risk. This dataset is global in extent and available at 0.5° x 0.5° latitude/longitude resolution. Other scPDSI datasets are available, including those from the US drought monitor. However, the US drought monitor dataset does not include AK and HI and only extends back to 1980.

Using the historical data from 1900-2020 creates an inherent assumption that past drought reflects current drought frequency. Given climate change, this assumption is biased but generally acceptable for two reasons:

We also calculate a bias-corrected future severe drought using the multimodal mean scPDSI from 25 CMIP6 models from Zhao and Dai (2022). This dataset is provided at 2.5° x 2.5° resolution. CMIP6 is considered the "state of the art” for understanding future climate. However, global models have intrinsic biases that add uncertainty to the projection. Specifically, the multi-model mean has much less variability and, thus, less severe drought months than the observations for the same historical period (Figure 2). In other words, the observed data indicates a greater range of drought than what the muti-model mean predicts.

Comparison of Modeled and Observed Drought illustrates the issue of the biases comparing the temporal distributions of scPDSI in the CMIP6 multi-model mean between 1900-2012 (end of the historical period) with the observed drought scPDSI between 1900-2020 for each grid cell. The green line in both plots indicates the threshold for severe drought (an scPDSI < -3). Each curve in the plot represents the cumulative distribution of the scPDSI for a given pixel. Thus, the value of the curve where it intersects the green line tells you the probability of a severe drought in a given year. Based on the historical observations, the CMIP6 data exhibits a range that is compressed more than expected. For example, in the CMIP6 model, most of the modeled historical distributions are close to 0 at the green line (meaning that the model estimates very little severe drought). For the actual historical record, the chance of severe drought ranges from 0.05 to approximately 0.6 (or more than half the time) for some areas. These plots indicate that CMIP6 produces low drought estimates relative to actual observed historical drought. While the UrbanFootprint dataset attempts to correct for the bias in CMIP6, bias correction is imperfect.

There are a few reasons why the CMIP6 multi-model mean has a narrower range. First, the model is at a coarser resolution, so areas that may be very dry or wet are getting averaged with the surrounding environments. Secondly, the multi-model mean reduces the variance within any given model.

Figure 1. Comparison of Modeled and Observed Drought

The cumulative distribution function of historical scPDSI at each grid cell in the CMIP6 multimodal mean (left) and CRU scPDSI dataset (right). The lack of extremes in the multimodal mean is apparent.

However, the CMIP6 multimodal mean does show meaningful change in severe drought risk. To preserve the higher spatial resolution (from CRU scPDSI) and reduce bias from the CMIP6 mean, we calculate the future change in months per decade of severe drought between the future scenarios and historical simulation in each grid cell and add that difference to our observation-based estimate of severe drought. This approach assumes that the models, specifically the multi-model mean, are better at predicting change than the absolute values. This assumption is widely supported in the literature, and bias correction is common practice (Thrasher et al., (2023), Larson et al. (2017), Forster et al., (2013), Navarro-Racines et al. (2020)).

We use 30-year windows to calculate the mean for a given future time period. For example, if over 30 years there were a total of 30 months of severe drought, then the reported value would be 10 months per decade. Months per decade here are units and do not correspond to the number in any given 10-year period. We calculate the drought frequency for the following scenarios and time periods.

Scenarios

Timeframes

Figure 2. Spatial Histograms of Observed, CMIP6 Modeled, and Bias-Corrected Future Drought

Spatial histograms of observation-based current (blue), CMIP6 mult-model future (orange), and bias-corrected future severe drought months per decade.

As described in the Hazard Summaries we intersect our gridded drought metrics with the UrbanFootprint parcel canvas and assign the highest intersect risk to that parcel. So, if a parcel spans two grid cells with 11 and 14 months per decade of severe drought, the parcel will be assigned a value of 14. We also calculate the area-weighted mean hazard value at census geographies such as county, tract, block group, or others. For example, suppose a county intersects three hazard grids. In that case, the mean value for the county will be calculated by summing the hazard values multiplied by their intersection area and dividing by the total county area.

Furthermore, we aggregate our parcel estimates of severe drought to census geographies to estimate exposure above a threshold or aggregation bin. For example, we would sum the population of all parcels in a county with severe drought frequencies of 36+ months per decade to give us the total population exposed to that threshold. Thus, we provide data for: