Methods for Mapping Impervious Surfaces An Exploratory Case Study from the Bassett Creek Watershed Josh Dunsmoor and Lucas Winzenburg December 10th, 2012
Outline What are Impervious Surfaces? Our Problem Approach Methods and Analysis Conclusion/Discussion
Impervious Surfaces “…any material of natural or anthropogenic source that prevents the infiltration of water into soil, thereby changing the flow dynamics, sedimentation load, and pollution profile of storm water runoff.” (Tilley, USGS) Rooftops, Roads/Streets/Highways, Sidewalks, Driveways, Parking Lots, Pools/Patios
Impervious Surfaces – Importance Fish populations Ground Water Assessment Flood Effects Heat Islands Studying Habitat Fragmentation Water Quality Assessment
Our Problem Tasked with estimating Leaf Biomass in priority watersheds in Minnetonka, MN Used to determine nutrient loading, specifically amount of pollutants from Leaf Litter (phosphorus, nitrogen) Where are the conveyances? What is the best method of mapping these conveyances?
Factors in Determining our Methods Scale/Study area size – approximately 1 sq. mile study area (existing data is 30m resolution) Data availability, cost Data quality (i.e. Leaf-on vs. Leaf Off, how recent was the capture?) Time constraints Lack of Experience Technology Limitations
Methods/Data Data Used Method Heads up Digitization Unsupervised Classification Supervised Classification LiDAR/HUD Hybrid OpenStreetMap .5m Aerial photo 2008 CIR NAIP Imagery (Leaf-on) State LiDAR data and manual digitization .osm extract converted into feature classes
Study Area
Heads-Up Digitization “Manual digitization by tracing a mouse over features displayed on a computer monitor, used as a method of vectorizing raster data.” (ESRI) Used ArcGIS 10.1 Pros – Accurate Cons – Labor Intensive, Costly (Time Consuming)
Heads-Up Digitization Imagery Final Classification
Heads-Up Digitization Class Breakdown Total % % Pervious/Impervious Pervious Ground 54% 64% Open Water 10% Buildings 9% 36% Streets 27% Accuracy Assessment Reference Data Map Data Total Pervious Ground Open Water Buildings Streets Producer's Accuracy 95 100.00% 17 28 6 21 1 75.00% 60 5 55 91.67% 200 106 56 User's Accuracy 89.62% 98.21% Total Accuracy 94%
Unsupervised Classification 2008 CIR NAIP Imagery- 1m, leaf-on Used ERDAS Imagine – ISODATA Algorithm 4 Attempts – Used varying amount of classes (4, 6, 10). More classes made classes less distinguishable. Deliverable used 3 classes, 25 iterations
Unsupervised Classification Imagery Final Classification
Unsupervised Classification Class Breakdown Total % % Pervious/Impervious Pervious Ground 38% 66% Open Water 28% Impervious 34% Reference Data Map Data Total Pervious Ground Open Water Impervious Surface Producer's Accuracy 82 67 11 4 81.71% 100.00% 57 5 7 45 78.95% 150 72 29 49 User's Accuracy 93.06% 37.93% 91.84% Total Accuracy 82%
Supervised Classification 2008 CIR NAIP Imagery – 1m, leaf-on Used ERDAS Imagine 4 attempts – used a varying amount of training areas and Max Likelihood and Minimum Distance to Mean Deliverable was 3 classes – Impervious, Vegetation, Water 10 Training Areas per class Minimum Distance to Mean
Supervised Classification Imagery Final Classification
Supervised Classification Class Breakdown Total % % Pervious/Impervious Pervious Ground 46% 67% Open Water 21% Impervious 33% Reference Data Map Data Total Pervious Ground Open Water Impervious Surface Producer's Accuracy 86 69 10 7 80.23% 16 2 14 87.50% 48 8 32 66.67% 150 79 39 User's Accuracy 87.34% 43.75% 82.05% Total Accuracy 77%
LiDAR/HUD Hybrid Obtained LiDAR data from MN Geospatial Information Office Used ArcMap for analysis LAStools third-party toolbox for data management Python scripting language for geoprocessing Initially tried to use LiDAR elevation/intensity only Settled on extracting buildings and using hybrid method
LiDAR/HUD Hybrid Elevation Points Intensity
LiDAR/HUD Hybrid Extraction of Buildings Python Script
LiDAR/HUD Hybrid Final Classification
LiDAR/HUD Hybrid Class Breakdown Total % % Pervious/Impervious Total % % Pervious/Impervious Pervious Ground 54% 65% Open Water 11% Buildings 9% 35% Streets 26% Reference Data Map Data Total Pervious Ground Open Water Buildings Streets Producer's Accuracy 116 108 1 7 93% 15 100% 20 6 14 70% 49 5 44 90% 200 119 16 51 User's Accuracy 91% 94% 86% Total Accuracy
OpenStreetMap Crowd-sourced, openly-editable online map Extracted underlying data directly from map Used ArcMap to convert into feature classes in GDB Buffered roads to 15ft, deleted extraneous layers Merged/clipped all impervious surface classes Areas not digitized as impervious/water were considered pervious
OpenStreetMap Web Map Final Classification
OpenStreetMap Class Breakdown Total % % Pervious/Impervious Total % % Pervious/Impervious Pervious Ground 62% 72% Open Water 10% Impervious 28% Reference Data Map Data Total Pervious Ground Open Water Impervious Surface Producer's Accuracy 83 78 5 94% 15 100% 52 13 39 75% 150 91 44 User's Accuracy 86% 89% Total Accuracy 88%
Overall Comparison of Methods
Overall Comparison of Methods
Pros/Cons of Methods Digitization Supervised Classification Accurate, Time Consuming for large areas Supervised Classification Quick, At mercy of imagery Unsupervised Classification Quicker, At mercy of imagery and algorithms LiDAR/HUD Hybrid Statewide coverage, Not a fully developed method, large dataset OpenStreetMap Fast/free, Sparse/imprecise, generalized data
Conclusions Heads-Up digitizzzzzzzz Leaf-off! Too many shadows with leaf-on Difficult to map at large scales due to shadows No “one size fits all” solution for RS problems Would likely choose LiDAR hybrid to perform task as it relates to our problem w/ this data
Further Research Object-based classification Explore LiDAR with greater depth High-res satellite data to improve accuracy?
Questions?