# UCRI California CA_2026.10: method specification Canopulse, 2026. This document describes how the released California Urban Canopy Risk Index (UCRI) values in `ucri_ca_2026_10_tracts.csv` were produced. Every rule below was confirmed against Canopulse's release-building code, the pipeline code that produced the component scores, the release metadata (`meta.json`, `MODEL_REGISTRY.json` and `SOURCE_MANIFEST.json`, published with the release data, for example https://www.canopulse.com/data/ca/current/meta.json) and the released values themselves. Where a detail could not be confirmed, this document says so. What is published and what is not: * Published in this package: the scoring logic, from tract input tables to component percentiles, UCRI, classes and ranks (`compute_ucri.py`); the weighting-sensitivity checks (`robustness/ucri_robustness.py`); and a check of the released arithmetic that runs on this package alone (`reproduce.py`). * Not published: the raster pipeline that turns the source rasters into tract summaries of tree canopy, impervious surface and land-surface temperature. ## 1. Release | Item | Value | |---|---| | Release | CA_2026.10 ("October 2026"), frozen 2026-10-08 | | Geography | 2020 census tracts, California: 9,129 records | | Fully scored | 9,041 (score kind `frozen`) | | Provisional | 66 | | Unscored | 22 | | Formula | UCRI = 0.25 CDS + 0.25 HES + 0.15 ISS + 0.20 SVS + 0.15 EBS | ## 2. Inputs, observation periods and caveats | Component | Raw tract value | Source and vintage | |---|---|---| | CDS, canopy deficit | 100 minus mean tree canopy cover (%) | NLCD Tree Canopy Cover 2025, USFS v2025-6 (NLCD-processed), USDA Forest Service | | HES, surface heat | Mean summer land-surface temperature (°F) | Landsat 8/9 Collection 2 Level-2 ST_B10, per-pixel mean of every pass under 20% scene cloud, 2026-06-01 to 2026-09-15 (424 passes) | | ISS, impervious surface | Mean fractional impervious surface (%) | Annual NLCD Collection 1.2 Fractional Impervious Surface, 2025 (USGS) | | SVS, social vulnerability | CDC/ATSDR SVI overall percentile `RPL_THEMES` (0 to 1) | CDC/ATSDR SVI 2022, California state file | | EBS, pollution burden | CalEnviroScreen 5.0 Pollution Burden percentile, field `PollutionP` (0 to 100) | OEHHA CalEnviroScreen 5.0, released 2026-07-01, CSV SHA-256 `3abb90971997cc151afe8c0ba46a01f544d1f8aaf4bc8b5d1f72c6d46007a296` | | Geography | Tract polygons, land and water area | TIGER/Line 2020 tracts, counties, places | **Canopy note (CDS row).** Tree canopy comes from a 30 m satellite model (NLCD Tree Canopy Cover 2025). It misses many street trees and small canopy in dense urban areas, so values are usually lower than a city's own 1 m canopy assessment. Use it to compare tracts, not as a local canopy inventory. **Surface heat note (HES row).** Surface heat is the daytime temperature of roofs, pavement and ground measured by Landsat satellites on summer mid-mornings (about 10 to 10:30 a.m. local time). It runs much hotter than air temperature and does not measure air temperature or nighttime heat. Tract summaries for the raster layers: each raster is placed on one EPSG:3857 grid of about 37.6 m (nearest neighbor for canopy and impervious, bilinear for temperature). The tract value is the mean of valid pixels whose centers fall inside the tract polygon (`all_touched=False`). Valid ranges: 0 to 100 for canopy and impervious; above 30 °F and below 200 °F for temperature. Landsat pixels are masked with QA_PIXEL bits 0 to 5 (fill, dilated cloud, cirrus, cloud, shadow, snow). SVI values of -999 are treated as missing. ### Observation periods Each input describes its own period: * Tree canopy: the 2025 layer of NLCD Tree Canopy Cover (v2025-6). * Impervious surface: the 2025 layer of Annual NLCD Collection 1.2. * Land-surface temperature: Landsat passes from June 1 to September 15, 2026. * Social vulnerability: CDC/ATSDR SVI 2022, built from American Community Survey 2018 to 2022 five-year estimates. * Pollution burden: CalEnviroScreen 5.0, released July 1, 2026. Each of its indicators has its own data years, listed in OEHHA's CalEnviroScreen 5.0 documentation. * Geography: 2020 census tracts. UCRI therefore combines measurements from different periods. It is not a snapshot of a single year. ### Social vulnerability (SVS) SVS starts from `RPL_THEMES`, the overall percentile ranking in the CDC/ATSDR SVI 2022 California state file. CDC builds it from 16 American Community Survey variables grouped in four themes: socioeconomic status; household characteristics; racial and ethnic minority status; and housing type and transportation. `RPL_THEMES` ranks tracts by the sum of all four theme sums, so it includes the racial and ethnic minority status theme. Canopulse uses the overall value as CDC publishes it and does not reweight, drop or add themes. In the state file, tracts are ranked within California. ### Land-surface temperature compositing (HES) Every Landsat 8 and Landsat 9 Collection 2 Level-2 scene over California from June 1 to September 15, 2026 with under 20% scene cloud cover was used: 424 passes over 54 Landsat path and row footprints, an average of about 8 passes per footprint. Passes were read from Microsoft Planetary Computer (collection `landsat-c2-l2`). Fill, dilated cloud, cirrus, cloud, cloud shadow and snow pixels were removed with QA_PIXEL bits 0 to 5. Each remaining ST_B10 value was converted to kelvin (DN × 0.00341802 + 149.0) and then to °F, and each pixel of the composite is the mean of all its valid passes, so the result does not depend on the order of the passes. The number of clear passes at a pixel varies with cloud and fog and with the overlap between neighboring paths. Landsat's thermal band has a native resolution of 100 m; the Level-2 product delivers it on a 30 m grid, and Canopulse resampled it bilinearly to the common grid above. Neither step adds spatial detail. ### Heat layer reliability To measure how much tract values depend on which days were observed, the passes in each path and row were alternated by date into two halves, and each half was composited and summarized by tract on its own. The statistics, as recorded in `SOURCE_MANIFEST.json` (`heat.reliability`): * Split-half Spearman correlation of tract values between the two halves, statewide: 0.90 (0.897). * Spearman-Brown estimate for the full composite, statewide: 0.95 (0.946). This is the figure quoted elsewhere as "rank reliability 0.95". * Split-half Spearman in the counties with at least 200 tracts where it is below 0.90: Alameda 0.80, Contra Costa 0.86 (0.864), Los Angeles 0.88 (0.875). Summer fog near the coast is the likely reason surface temperature varies more from pass to pass there. What "rank reliability 0.95" means: it estimates how closely the order of tracts in the full 424-pass composite would agree (Spearman correlation) with the order from another composite built from the same number of comparable passes. A value of 1 would mean the same order. It measures how repeatable the tract ranking is across observation days. It does not show that land-surface temperature represents air temperature or heat risk. Canopulse fixed targets for the split-half statistic before the run finished (0.95 statewide and 0.90 in every county with at least 200 tracts); the layer was released below both. These figures are Canopulse's own check, not an independent one. ## 3. Percentile normalization and reference populations **CDS, HES, ISS and SVS** are statewide percentile ranks computed by Canopulse: ``` percentile = average_rank(value) / n * 100 ``` * Implemented in the pipeline as pandas `Series.rank(pct=True, method="average") * 100`; `compute_ucri.py` does the same in standard-library Python and gives identical values. * Ranks are ascending, so a higher raw value (less canopy, hotter, more impervious, more vulnerable) gives a higher score. * Ties receive the average of the ranks they span. * `n` is the number of tracts with a non-missing raw value for that component, after excluding the 22 water-only tracts (land area 0 and water area above 0). In this release n = 9,107 for CDS, HES and ISS (the 9,041 fully scored tracts and the 66 provisional tracts) and n = 9,041 for SVS. The reference population is therefore all land tracts with that layer: not only the 9,041 fully scored tracts, and not all 9,129 records. * Before rounding, the lowest value receives 100/n (about 0.011 here), so the lowest published values round to 0.0 (4 tracts in each of CDS, HES, ISS and SVS). An untied maximum receives 100. Tied maxima receive the average of the ranks they span, so the published CDS maximum is 99.8 (34 tracts with zero canopy). * Each percentile is rounded to one decimal before it is used (pandas `round`: the value times 10 is rounded half to even, then divided by 10). | Component | Ranked by | Reference population | |---|---|---| | CDS, HES, ISS | Canopulse | 9,107 land tracts with the layer | | SVS | Canopulse | 9,041 land tracts with a CDC/ATSDR SVI value | | EBS | OEHHA (used as published, not re-ranked) | OEHHA's CalEnviroScreen 5.0 tract set; 9,106 California tracts matched | `compute_ucri.py`, run by Canopulse on the pipeline's unrounded tract tables (not included in this package), reproduces every released value in the CSV exactly: 9,107 of 9,107 for CDS, HES and ISS, 9,041 of 9,041 for SVS, and all EBS values, scores, classes and ranks. **SVS** re-ranks the CDC SVI overall percentile within the California tracts that have it. CDC's own `RPL_THEMES` is already a percentile ranking among California tracts, so this step is a monotone rescaling, not a new measure. The released SVI value (`cdc_svi_rpl_themes` in the CSV) is rounded to 0.1 on the 0 to 1 scale, so SVS cannot be recomputed from the CSV alone. **EBS** is not ranked by Canopulse. It is OEHHA's own published statewide `PollutionP` percentile, rounded half up to one decimal, joined on the 2020 tract GEOID (9,106 of 9,129 tracts match; 23 have no CalEnviroScreen 5.0 row). Canopulse has not independently re-derived how OEHHA computes `PollutionP`; see the CalEnviroScreen 5.0 documentation. Its reference population is OEHHA's tract set, not the Canopulse set above. **Not confirmed:** whether OEHHA's percentile uses the same tie and endpoint convention as the Canopulse percentiles. It does not affect reproduction of UCRI from the released components, but it means the five components are percentiles over slightly different populations and conventions. ## 4. Weights and score ``` UCRI = 0.25*CDS + 0.25*HES + 0.15*ISS + 0.20*SVS + 0.15*EBS ``` The release builder computes the weighted sum from the one-decimal component values, adding terms left to right in the order CDS, HES, ISS, SVS, EBS in IEEE double precision, and then rounds half up to one decimal: `Math.round((x + Number.EPSILON) * 10) / 10`. Note for reimplementers: Python 3.12 and later use compensated summation in `sum()` for floats. That changes the last bits of about 0.6% of sums that sit exactly on a .x5 boundary and moves them by 0.1 after rounding. Add terms in a plain loop (as `reproduce.py` and `compute_ucri.py` do) to match the release exactly. ## 5. Score kinds and the provisional estimate | Kind | Rule | Count | In this release | |---|---|---|---| | Fully scored (`frozen`) | All five components present | 9,041 | Ranked and classified | | Provisional | One to four components present | 66 | 65 missing SVS only; 1 (06075980401) missing SVS and EBS | | Unscored (`none`) | No component present | 22 | Exactly the 22 water-only tracts | **Provisional estimate:** the sum of available weighted contributions divided by the sum of available weights, rounded half up to one decimal: ``` ucri_provisional = round( sum_available(w_k * c_k) / sum_available(w_k), 1 ) ``` This equals renormalizing the available weights to sum to 1. Verified on all 66 provisional records. Provisional estimates are excluded from ranks, priority classes and thresholds. Unscored tracts still carry raw physical values in the CSV (for example canopy and temperature over water), but no component scores. ## 6. Priority classes Priority classes are fifths (quintiles) of the 9,041 fully scored tracts, ranked by UCRI. The cut points are the 20th, 40th, 60th and 80th percentiles of the 9,041 released UCRI values (linear interpolation at index (n-1) × q, the numpy and pandas default), each rounded half up to one decimal. They are stored in the release metadata (`meta.json`, `priority_thresholds`: Low 30, Moderate 44, High 56.9, Extreme 70.3). A tract takes the highest class whose cut point its UCRI meets or exceeds. Provisional and unscored tracts have no class. | Stored value | Displayed on the website and in reports | UCRI range | Tracts | |---|---|---|---| | Very Low | Very low screening priority (bottom fifth of California tracts) | below 30.0 | 1,799 | | Low | Low screening priority (fourth fifth of California tracts) | 30.0 to below 44.0 | 1,804 | | Moderate | Moderate screening priority (middle fifth of California tracts) | 44.0 to below 56.9 | 1,808 | | High | High screening priority (second fifth of California tracts) | 56.9 to below 70.3 | 1,816 | | Extreme | Very high screening priority (top fifth of California tracts) | 70.3 and above | 1,814 | The stored class value "Extreme" is displayed as "Very high screening priority" (top fifth statewide) on the Canopulse website and in its reports. The CSV and the release data keep the stored value. The classes are relative screening categories, not hazard levels. UCRI classes are not official designations. They do not show SB 535 disadvantaged community status, AB 1550 low-income community status or eligibility for any program. Check the official CalEPA and California Climate Investments maps for those. Official maps: CalEPA SB 535 disadvantaged communities (2026 designation), https://calepa.ca.gov/programs/dac2026/; California Climate Investments priority populations maps, https://www.caclimateinvestments.ca.gov/priority-populations. Class sizes differ slightly from 20% because many tracts share a score at one decimal. ## 7. Ranks `state_rank` runs 1 to 9,041 over fully scored tracts, by UCRI descending, with ties broken by GEOID ascending. County and place ranks (in the release data, not in this CSV) use the same rule within each county and within each place; a place that spans two counties is ranked separately in each. The release data also carry a position percentile within each group, (rank minus 1) / (n minus 1) × 100, rounded half up to one decimal, so the top-ranked tract has 0. ## 8. Confidence label Each tract record in the release data carries a confidence label (it is not a column in this CSV). In this release: | Label | Tracts | |---|---| | High | 0 | | Moderate | 9,041 (every fully scored tract) | | Low | 0 | | Insufficient | 88 (66 provisional and 22 unscored) | Every fully scored tract is Moderate because the label describes the input data, which are the same kind for every tract, not the tract itself. High requires every contributing layer at a native resolution of 30 m or finer with a vintage within two years. Three inputs trigger the Moderate rule for every tract: Landsat land-surface temperature is natively 100 m (resampling it for alignment adds no detail), and CDC/ATSDR SVI and CalEnviroScreen Pollution Burden are tract-level composite indices. The label therefore does not tell one fully scored tract from another. Tracts without a full score are Insufficient. ## 9. Rounding summary | Value | Rounding | |---|---| | Raw canopy, impervious, temperature | 0.1 (pipeline, pandas `round`) | | CDC SVI `RPL_THEMES` as released | 0.1 on a 0 to 1 scale | | CDS, HES, ISS, SVS | 0.1 (pipeline, pandas `round`) | | EBS | 0.1, half up | | UCRI and provisional estimate | 0.1, half up, from the rounded components | | Thresholds | 0.1, half up | ## 10. CSV columns | Column | Meaning | |---|---| | `geoid` | 2020 tract GEOID (11 digits) | | `county`, `place` | County name; place by largest land-area overlap (under 25% overlap is reported as unincorporated in the release) | | `score_kind` | `frozen` (fully scored), `provisional` or `none` | | `cds` ... `ebs` | Component percentiles 0 to 100; blank when missing | | `ucri` | Released UCRI; blank unless fully scored | | `ucri_provisional` | Provisional estimate; blank unless provisional | | `priority`, `state_rank` | Stored class (`Extreme` is displayed as "Very high screening priority") and statewide rank; blank unless fully scored | | `canopy_pct` | Tract mean tree canopy cover, % (see the note below) | | `summer_lst_f` | Tract mean summer land-surface temperature, °F (see the note below) | | `impervious_pct` | Tract mean impervious surface, % | | `cdc_svi_rpl_themes` | CDC SVI overall percentile, 0 to 1, rounded to 0.1 | | `ces5_pollution_burden_pctl` | CalEnviroScreen 5.0 `PollutionP` as used for EBS | `canopy_pct`: Tree canopy comes from a 30 m satellite model (NLCD Tree Canopy Cover 2025). It misses many street trees and small canopy in dense urban areas, so values are usually lower than a city's own 1 m canopy assessment. Use it to compare tracts, not as a local canopy inventory. `summer_lst_f`: Surface heat is the daytime temperature of roofs, pavement and ground measured by Landsat satellites on summer mid-mornings (about 10 to 10:30 a.m. local time). It runs much hotter than air temperature and does not measure air temperature or nighttime heat. ## 11. Limitations * **Surface heat is not air temperature.** HES and `summer_lst_f` measure the daytime temperature of roofs, pavement and ground at the late-morning Landsat pass. They do not measure air temperature or nighttime heat. * **Statewide percentiles rank hot, low-canopy regions high by construction.** Every component is ranked against all California tracts, so desert and Central Valley tracts, which are hot and naturally have little tree canopy, score high on canopy deficit and surface heat because of their climate and biome. CDS answers "how little canopy does this tract have compared with all California tracts", not "how far is it below the canopy it could achieve". County and place ranks in the release data give a within-area view. * **Tract averages describe tracts, not residents.** Each value is an average over a whole tract. It does not describe every block or every person in the tract (the ecological fallacy). * **Tract scale.** Tracts differ widely in size and population: some rural and desert tracts are very large, and some tracts have very few residents. Canopulse flags tracts with fewer than 500 residents (ACS 2018 to 2022 estimate, as carried in CDC/ATSDR SVI 2022) or a Census tract code from 9800 to 9899 (special land-use tracts such as industrial areas, airports or parks). In this release 34 fully scored tracts meet one of these conditions, 6 of them in the top class. Flags never change a score or rank. * **Classes are statewide fifths.** 36 of California's 58 counties have no tract in the top class, because the classes divide all fully scored tracts statewide, not the tracts of each county. * **The weights are a judgment.** They were not estimated from outcome data, and UCRI has not been tested against heat illness, mortality or other outcomes. `robustness/REPORT.md` tests sensitivity to the weights only. ## 12. How to reproduce Requirements: Python 3.8 or later. No packages. ``` python3 reproduce.py ``` or, to check a copy elsewhere, `python3 reproduce.py path/to/ucri_ca_2026_10_tracts.csv`. The script: 1. Recomputes UCRI for the 9,041 fully scored tracts from the five component columns and reports the maximum absolute difference and the number of mismatches beyond 0.05. 2. Recomputes the 66 provisional estimates. 3. Checks every score kind against the availability rule. 4. Recomputes the quintile thresholds and every priority class. 5. Recomputes every statewide rank. 6. Approximately checks the CDS, HES and ISS percentile step from the rounded raw values (exact agreement is not expected because rounding the raw values creates extra ties). It exits with code 0 only if steps 1 to 5 match the release. Output recorded by Canopulse on 2026-10-11 with Python 3.14.7, from `python -I reproduce.py` run in the package folder (verbatim): ``` Read 9,129 tract rows from ucri_ca_2026_10_tracts.csv 1. UCRI, fully scored tracts tracts: 9,041 max |unrounded weighted sum - released UCRI|: 0.0500 recomputed and rounded to 0.1, identical to released: 9,041 of 9,041 mismatches beyond 0.05: 0 2. Provisional estimates (sum of available contributions / available weight) tracts: 66 max |unrounded estimate - released|: 0.0500 mismatches beyond 0.05: 0 3. Score kind from component availability counts: fully scored 9,041, provisional 66, unscored 22 rows whose score kind disagrees with the rule: 0 4. Priority thresholds (quintiles of released UCRI, linear interpolation, rounded to 0.1) recomputed: {'Low': 30.0, 'Moderate': 44.0, 'High': 56.9, 'Extreme': 70.3} published: {'Low': 30.0, 'Moderate': 44.0, 'High': 56.9, 'Extreme': 70.3} priority classes that disagree: 0 5. Statewide rank (UCRI descending, ties by GEOID ascending) ranks that disagree: 0 6. Percentile step, approximate (raw values in the CSV are rounded, which creates extra ties) CDS: n=9,107, median |diff| 0.14, max 0.73, within 0.5 points 95.8% HES: n=9,107, median |diff| 0.08, max 0.44, within 0.5 points 100.0% ISS: n=9,107, median |diff| 0.04, max 0.25, within 0.5 points 100.0% SVS is not checked here: the released SVI value is rounded to 0.1 on a 0 to 1 scale. EBS is the official CalEnviroScreen 5.0 PollutionP percentile, not re-ranked by Canopulse. RESULT: PASS, released scores reproduce exactly ``` Verify file integrity first with `sha256sum -c SHA256SUMS.txt` (Linux, macOS with coreutils) or `Get-FileHash` on Windows. The scoring step: `compute_ucri.py` (Python 3.8 or later, no packages) needs tract input tables that are not in this package: the tree canopy, impervious surface and land-surface temperature tract summaries from the raster pipeline, plus the tract list, the CDC/ATSDR SVI 2022 California file and the CalEnviroScreen 5.0 CSV (the last three are public). Its header lists the columns it reads. ``` python3 compute_ucri.py --tracts tracts.csv --canopy canopy.csv \ --impervious impervious.csv --lst lst.csv --svi SVI_2022_California.csv \ --ces calenviroscreen50.csv --places places.csv \ --out ucri_computed.csv --check ucri_ca_2026_10_tracts.csv ``` With `--check` it compares its result with the released CSV column by column and reports `RESULT: PASS` only if every value matches. Canopulse ran it this way on 2026-10-11 on the pipeline's tract tables: every value matched, and the county and place ranks it computes also matched the release data. The weighting-sensitivity checks: `python3 robustness/ucri_robustness.py --out rerun` (Python 3.8 or later, no packages, about a minute). See `robustness/REPORT.md`. What this reproduces: the arithmetic from released components to released scores, classes and ranks (`reproduce.py`, on this package alone), and from tract input tables to components, scores, classes and ranks (`compute_ucri.py`, given those tables). What it does not reproduce: the raster processing and tract summaries, which need the source rasters (tens of GB) and the raster pipeline code, which is not published.