Mapping urban form into local climate zones for the continental US from 1986-2020

Abstract

Urbanization has altered land surface properties driving changes in micro-climates. Urban form influences people's activities, environmental exposures, and health. Developing detailed and unified longitudinal measures of urban form is essential to quantify these relationships. Local Climate Zones [LCZ] are a culturally-neutral urban form classification scheme. To date, longitudinal LCZ maps at large scales (i.e., national, continental, or global) are not available. We developed an approach to map LCZs for the continental US from 1986 to 2020 at 100 m spatial resolution. We developed lightweight contextual random forest models using a hybrid model development pipeline that leveraged crowdsourced and expert labeling and cloud-enabled modeling - an approach that could be generalized to other countries and continents. Our model achieved good performance: 0.76 overall accuracy (0.55-0.96 class-wise F1 scores). To our knowledge, this is the first high-resolution, longitudinal LCZ map for the continental US. Our work may be useful for a variety of fields including earth system science, urban planning, and public health.

Fig. 1

LCZ classification scheme including (a) 10 built classes (i.e., LCZs 1–10) and (b) 7 land cover classes (i.e., LCZs A-G). Panels (a) and (b) are adapted from Fig. 1 in Demuzere et al.; representative aerial maps (© Google Earth 2020) for each LCZ type are shown. (c) The spatial-temporal distribution of LCZ TAs across CONUS. The colors of TAs are consistent with the color palette presented in panels (a) and (b). (d) Counts of (1) total TAs by year (red line), (2) TAs by LCZ class in 2020 (black solid bars), and (3) sampled points within TA polygons by class for each year (black stripe bars) for model development.

Fig. 2

Model development workflow. (a) TA polygon labeling including crowdsourced labeling using the MTurk platform for the base year (2017) and manual labeling by experts for other years (1986–2020). (b) Balanced label sampling from TA polygons and model feature extraction including features collected from yearly Landsat composite layers, LULC layers (LCMAP and LCMS), and Census layers. (c) Model training, fine-tuning, and evaluation on a local machine using Python scikit-learn library. (d) Model transfer from the local machine to the GEE platform, post-classification processing, and final LCZ prediction surfaces at 100 m × 100 m resolution from 1986–2020 across CONUS.

Fig. 3

Model performance based on the 5-fold spatial cross validation. (a) Overall model performance by the local model, GEE model, and final model (i.e., GEE model with spatial-temporal post-classification processing). Reported metrics include the overall accuracy (OA), overall accuracy for urban classes (OA_u), overall accuracy for built versus natural classes (OA_bu), and weighted accuracy (OA_w). (b) Class-wise model performance by the local model, GEE model, and final model. (c) Spatial distribution of the final model performance by state. Reported metrics include OA, OA_u, OA_bu, and OA_w.

Fig. 4

Confusion matrix for the final LCZ model. The value in each cell is the number of test samples. Precision is shown to the bottom of the confusion matrix. Recall and F1 scores are shown to the right of the confusion matrix.

Fig. 5

Temporal consistency for the final LCZ model. (a) Overall model performance from 1986 to 2020, including the overall accuracy (OA), overall accuracy for urban classes (OA_u), overall accuracy for built versus natural classes (OA_bu), and weighted accuracy (OA_w). (b,c) show the average yearly transition (unit: %) among LCZ categories from 1986 to 2020 for CONUS. (b) shows the transition for the GEE model with only spatial filtering. (c) shows the transition of our final LCZ model, i.e., the GEE model with both spatial and temporal filtering. The transition direction is from rows to columns. For example, in (c) row 1, 93% of LCZ 1 pixels in CONUS remained LCZ 1 from a previous year to the current year on average.

Fig. 6

LCZ mapping. (a) CONUS prediction surface at 100 m × 100 m resolution using year 2020 for illustration. (b) LCZ mapping for an example city (Las Vegas, Nevada) at 100 m × 100 m resolution by year. For illustration, change in urban form is shown based on a ~5-year interval and the Census 2020 Urban Areas boundary. Full results for all years (1986 to 2020) are available in our LCZ dataset.

Fig. 7

Thematic benchmarking. (a) The binary (urban vs. natural) LCZ and NLCD maps using 2019 for illustration. The threshold for defining urban vs. natural was 10% impervious surface fraction. LCZ 9 (sparsely built) was not considered in the thematic benchmarking because the impervious fraction of LCZ 9 pixels may fall into either category. The comparison was conducted for both CONUS, and CONUS-UA which used Census 2020 Urban Areas boundary. (b) Comparison results for overall accuracy and the urban and natural pixels.

Fig. 8

LCZ maps for 6 example metropolitan areas (two cities each for the western, central, and eastern US) in 1986 and 2020 with corresponding aerial maps from Google Earth (© Google Earth 2020). Blue lines in the aerial maps are the Census 2020 Urban Areas boundaries. Land and water area within the boundaries are listed for reference.

Fig. 9

Trends of LCZ composition ratio from 1986 to 2020 for 6 US metropolitan areas. All composition ratios were calculated based on the Census 2020 Urban Areas boundary shown in Fig. 8.

Fig. 10

The chord diagram to show the major transition of LCZ classes between year 1986 and year 2020 for 6 US metropolitan areas. The directions of the transitions are denoted by the arrows. The colors of the chord are consistent with the colors of the transition sources. The ticks on the arcs represent the transition rate. For example, as shown by the LCZ C arc in Denvor-Aurora, 7% of pixels transit from LCZ C to LCZ 6 (chord color consistent with LCZ C); 4% from LCZ C to LCZ F (chord color consistent with LCZ C); 0% from other LCZ types to LCZ C. To illustrate major transitions, any LCZ transition that occurred in less than 1% of city pixels were removed from the chord diagram; pixels remaining the same type were also removed. Full transition details can be found in Supplementary Information Fig. S4. All transition rates were calculated based on the Census 2020 Urban Areas boundary shown in Fig. 8.