Forecasting San Antonio Urban Growth to 2030 (Markov-CA)

Predicted land cover 2030 — Simulated 2030 land cover: urban growth concentrates near the existing core and major highways

01Why forecast growth

Urban growth reshapes watershed hydrology: more impervious surface means more runoff, more flood risk, and less groundwater recharge. Knowing where a city will expand lets planners act before the change happens.

The model produces a prediction, built from observed transitions and validated against what actually happened.

Delineated watershed study area — Watershed of San Antonio River near Falls City (~5,453 km²)

02Training period (2010→2020)

Delineate watershed from DEM (Fill → Flow Direction → Flow Accumulation → Snap → Watershed).
Reclassify NLCD 2010 & 2020 into 5 classes.
Compute the 2010→2020 Markov transition matrix in Python.
Build CA spatial factors: distance to roads, highways, facilities, plus 3×3 neighborhood influence.
Calibrate weights & threshold against observed 2020, then simulate 2030.

Calibration workflow — Iterative calibration loop to maximize Figure of Merit

03Transition probabilities

The Markov matrix quantifies how likely each land cover was to change between 2010 and 2020. Built land was highly stable (>99% stays built), while barren, agriculture, and forest fed the growth.

Neighborhood influence (β4 = 7) was the dominant driver: new growth clusters next to existing development.

From \ To	Built	Barren	Forest	Agri
Built	0.994	0.004	0.001	0.001
Barren	0.041	0.901	0.045	0.012
Forest	0.022	0.045	0.932	0.001
Agriculture	0.032	0.087	0.000	0.880

Neighborhood influence layer — Neighborhood influence, the strongest driver

04Validation and 2030 forecast

0.85Figure of Merit (validation)

7.6% / 6.6%False negatives / false positives

~5,453 km²Watershed study area

I compared the model's predicted 2020 against observed 2020. The strong match confirmed the model before I used it to forecast 2030. The 2030 simulation shows continued growth near the city core and along highways.

2020 validation — Predicted vs. observed 2020 (validation)

New growth forecast 2030 — Where new growth is forecast by 2030

Built-up area trend 2010-2030 — Built-up area share rising across 2010 → 2020 → 2030

05Code

The model is driven by two Python scripts, both on GitHub:

markov_transition_matrix.py reads the 2010 and 2020 land-cover rasters and computes the Markov transition count and probability matrices with NumPy and rasterio.
calibration_figure_of_merit.py compares the simulated 2020 built-up map against observed 2020 to compute the Figure of Merit, false positives, and false negatives (arcpy + NumPy).

06Limitations & next steps

Drivers limited to proximity and neighborhood (next: demographics, economics, policy, climate).
Two-date Markov assumes stationary transitions; 30 m resolution; single watershed.

Read the code

Python scripts (Markov matrix + calibration), README, and all maps on GitHub.

View on GitHub → Back to portfolio