Analysis of Multitemporal UAS Data and Its Applications

Center for Geospatial Analytics at North Carolina State University

Corey White adapted from Helena Mitasova, Anna Petrasova, and Justyna Jeziorska

Objectives

• Design a UAS monitoring campaign: metrics, spatial and temporal resolution, ground control
• Register a series of DSMs as a space-time raster dataset and query it by time
• Aggregate the series over time and compute per-cell statistics
• Derive core and envelope surfaces and use them to find distorted surveys
• Apply the series to crop height and biomass, viewsheds, and volumes

• 5A fused one UAS DSM with the lidar; today we work with all seven 2015 flights of Mid Pines as a time series
• The GRASS temporal framework does the bookkeeping: which map is from when, what they have in common, and how they change
• The lab (5B) registers the seven DSMs, finds the two distorted ones, and applies the series to crop biomass, a webcam viewshed, and a building volume
• Every concept today has a GRASS tool next to it; the lab uses them in the same order

Monitoring with UAS

• Section 1: what UAS monitoring is good for, and how to design a campaign so that the data can be analyzed as a series

UAS for monitoring change

Low cost and rapid deployment make UAS suited to monitoring change at the local scale: fields, small watersheds, construction sites

• Crop monitoring: growth, disease, stress
• Erosion: coastal, stream bank, rills and gullies
• Natural hazards: flooding, landslides, fire
• Industrial: mining, construction sites, stockpiles

• Airborne lidar is flown every few years; a UAS can fly the same field every two weeks
• The common thread is a surface that changes faster than any regional dataset is updated
• Mid Pines in 2015 is a crop monitoring series: seven flights from March to October

Monitoring design

Decision	What to settle before the first flight
Metrics	Relative height, volume, feature migration: what quantity answers the question
Spatial resolution	Fine enough to capture the change, coarse enough to process a series
Temporal resolution	Regular intervals, or flights timed to events such as storms or harvest
Accessibility	Flying over people, line of sight, places for GCPs
Georeferencing	The same GCPs surveyed once and used for every flight

• The metric decides everything else: crop height needs a DSM and a bare ground; a stockpile volume needs two DSMs
• The hardest requirement is the last one: a series is only comparable if every flight is referenced the same way
• The Mid Pines flights used different numbers of GCPs, and the lab finds out what that did to two of them

Processing a time series

• Interpolate every point cloud to a DSM on the same aligned grid at a common resolution
• Rasters suit a series better than TINs: the same cell in every map
• Check each DSM against GCPs and permanent features; correct shifts, flag distortions (5A)
• Give every DSM a time stamp and register the set in the temporal framework

• Aligned grid means the same extent, resolution, and cell origin, otherwise per-cell statistics compare different ground
• The quality check per flight is 5A’s difference map and profile
• Registering the maps is what turns a folder of rasters into a dataset you can query by time

Space-time datasets in GRASS

• Section 2: the temporal framework, time stamps, and the tools that list and plot a registered series

The temporal framework

Space-time raster dataset (STRDS): a set of raster maps registered in a temporal database, each with a time stamp; the maps are the states of a changing surface, and the dataset is treated as one object (Gebbert and Pebesma 2014)

• Efficient processing, management, and analysis of space-time data
• The dataset can hold rasters, 3D rasters, or vectors
• Tools work on the whole dataset: list by time, aggregate, compute statistics, apply map algebra
• In GRASS: t.create makes the dataset, t.register adds the maps with their times

• The maps themselves do not change; the database records which map belongs to which time
• Everything that follows, from aggregation to core and envelope, is a t.* tool reading that database
• Lab Part 1: t.create for uas_dsm, then t.register with a text file of map names and dates

Time stamps

Time stamp: the time assigned to one map in the dataset, as an instant (2015-06-20) or an interval (2015-06-01 to 2015-07-01), in absolute calendar time or in relative units such as days or years

• A UAS survey is a snapshot: an instant
• Aggregating snapshots over months or seasons turns them into intervals
• The dataset’s temporal type, absolute or relative, is set when it is created

• Instant versus interval is independent of absolute versus relative; the lab uses absolute instants, one date per flight
• The series file is two columns: map name and date, separated by a pipe
• t.info reports the type and the time span of the dataset after registration

The registered series

• The Timeline Tool shows when each map was taken and, in 3D, its spatial extent
• t.rast.list lists the maps with their times and statistics, optionally filtered with a where clause

• Seven flights: March, June, late July, mid August, two in September, early October
• The 3D plot shows the extents differ from flight to flight, which matters for the intersection in Section 3
• The lab lists maps with where="start_time >= '2015-08-01'"; SQL on the time column

Temporal plot

The Temporal Plot Tool plots the value of every map in the series at one location

• Pick a point on a road, in a field, and on a building and compare the three curves
• A field should rise through the season and drop at harvest; a road should stay flat
• The August spike at both points is a warning: that survey is off, which Section 3 confirms

Aggregation and per-cell statistics

• Section 3: what the series has in common, how to aggregate it, and the per-cell statistics that summarize change

Count and intersection

Count: maps with a value in each cell

Intersection: cells covered by every map

In GRASS: t.rast.series input=uas_dsm method=count output=count, and with -n the count is null where any map is null, which is the intersection

• The flights had different extents, so per-cell statistics are only meaningful where every map has data
• The intersection raster becomes the mask for the rest of the lab: r.mask raster=intersection
• Lab Part 2 computes both and zooms the region to the intersection

Temporal aggregation

Temporal aggregation: combining the maps that fall in each interval of a chosen granularity into one map with a statistic, such as the monthly mean; t.rast.aggregate

• September has two flights; a monthly aggregation averages them into one September map
• The output is a new dataset with interval time stamps, one per month that has data
• Lab Part 2: t.rast.aggregate ... granularity="1 months" method=average

Per-cell statistics over time

• Mean and standard deviation of the elevation at each cell
• Minimum, maximum, and range
• Time of the minimum and time of the maximum: which map it came from
• Linear regression through time: slope, offset, and coefficient of determination

In GRASS: t.rast.series input=uas_dsm method=minimum|maximum|min_raster|max_raster|average|stddev|slope ... (Mitasova et al. 2009)

• Each statistic is one raster computed down the stack of maps, cell by cell
• min_raster and max_raster return the index of the map, counted from zero, not the elevation
• The lab uses minimum, min_raster, and maximum; regression needs a longer, regular series

Core and envelope

Core surface: the minimum elevation measured at each cell over the series. Envelope: the maximum. Everything the surface ever did lies between them

Cutting planes through the lidar ground, the core, and the envelope of the series

• On a crop field the core is close to bare ground and the envelope is the canopy at its tallest
• On a building both are the roof, so the difference is zero: a built-in test of the series
• The two surfaces come from t.rast.series with method=minimum and method=maximum

Finding distorted surveys

Lidar ground minus core

Index of the map with the minimum

• Compute the core and subtract it from the lidar bare ground
• Where the core lies well below the ground, min_raster names the survey responsible

• A bare ground DSM cannot be below the real ground, so a core far under the lidar DEM is an error
• Here indexes 2 and 3, the late July and mid August flights, flown with two and three GCPs, supply most of those minima
• Lab Part 3: r.what on min_raster at two points, then exclude the two surveys with a where clause

Corrected core surface

Without the two distorted surveys

Histogram-equalized colors

• Remove the distorted DSMs and recompute the core: the difference to the lidar ground is now small
• A histogram-equalized color table stretches the small differences and reveals a shift between two lidar swaths

• The second map is the same data; only the colors changed, and a straight line appears
• That line is a seam between two lidar flight lines, a few centimeters apart: the corduroy effect from 5A
• Equalized colors are the way to find small systematic patterns in a difference map

Envelope and range

Lidar ground minus envelope: the tallest crop at each cell

The envelope surface in 3D

• Envelope minus ground: the maximum crop height over the season at each cell
• Time of maximum: when each cell was highest
• Range: where the surface varied most

• The envelope is the season’s tallest crop, field by field, in one map
• max_raster gives the flight of the maximum, which differs between fields planted at different times
• The lab uses the envelope for the webcam viewshed, as the worst case for visibility

Envelope of a parking area

• The envelope shows every car that was parked at the site during any flight
• Core, snapshot, and envelope surfaces together describe how an area is used

• A light example of the same statistic: anything that was ever there is in the envelope
• The core, by contrast, shows the lot empty
• Lab Part 5 asks you to look at the relief of the envelope DSM around the building for exactly this

Regression through time

• Needs a well designed, systematic series: regular flights over a longer period
• Choose the part of the series where the change is close to linear, such as a growth period
• Per-cell linear regression gives a map of the rate of change and a map of the fit

• Seven irregular flights with two bad ones are not enough for a regression; this is where a designed campaign pays
• The slope map is the growth rate in meters per day, cell by cell
• t.rast.series method=slope and method=detcoeff

Map algebra on a series

• Apply one map algebra expression to every map in the series, cell by cell
• The result is a new series, registered as a new dataset with the same time stamps
• Simpler than temporal algebra, which combines maps across time

In GRASS: t.rast.mapcalc inputs=uas_dsm expression="uas_dsm - lidar_dem" output=height basename=height, where uas_dsm stands for each map in turn

• The dataset name in the expression is replaced by each registered map
• Other rasters in the expression, like the lidar DEM, are the same for every map
• Lab Part 4 uses this to turn the DSM series into a crop height series

Applications

• Section 4: three uses of the series the lab works through: crop height and biomass, a viewshed, and a volume

Crop height series

June 20

July 27

October 6

Height above the lidar bare ground between 0.3 and 2 m, from t.rast.mapcalc with if() on the fields raster

• The window of 0.3 to 2 m keeps the crop and drops the soil noise below and the building and trees above
• The July map shows distortion rather than crop: the pattern ignores the field boundaries
• Lab Part 4: t.rast.mapcalc with the fields mask, then t.rast.univar for the statistics per date

Crop volume

• Sum of the crop height over the field cells, times the cell area, is the crop volume above ground
• A proxy for above ground biomass, and its change through the season

• r.univar gives the sum of heights; multiply by the cell area for cubic meters
• Comparing the volume per flight gives the growth curve of the field
• Lab Part 4 computes the June volume from veg_uas_lidar_2 with r.univar

Volume of a structure

• Difference between the DSM and the bare ground over the footprint of the feature
• r.volume sums the difference over a clump of cells and reports the volume

• Same idea as the crop volume, for one object: a building, a stockpile, a dune
• The footprint comes from a threshold on the difference, here everything more than 2 m above the ground
• Lab Part 6 does this for the building at Mid Pines

Viewsheds on a changing surface

Cumulative viewshed: the number of observer points from which each cell is visible; r.viewshed per observer, then r.series method=count

• The vegetation height decides what a camera sees, so the viewshed changes through the season
• Compute the viewshed on the lidar DSM and on the envelope DSM to see the best and worst case

• Different colors are the viewshed extent at different times: the crop closes the view as it grows
• Lab Part 5 sites a webcam: a cumulative viewshed from six candidate points on the lidar DSM, then the view from the chosen point on the envelope DSM
• The envelope DSM is built by fusing the UAS envelope into the lidar DSM with r.patch.smooth from 5A

Visualizing a series

• Animation Tool: one frame per map, draped over a shaded relief, exported as a GIF (5A)
• Temporal Plot Tool: the value through time at a point
• 3D view of two surfaces with a cutting plane: core, envelope, and lidar ground
• The GRASS temporal workshop has the terrain time series example

• The flow animation in 5A used the same Animation Tool on a series of depth maps
• For a DSM series, drape the crop height over the relief and step through the flights
• The notebook does this with a series map in place of the GUI tool

Key terms

Space-time raster dataset: a set of rasters registered with time stamps and handled as one object

Time stamp: an instant or an interval, in absolute or relative time, assigned to each map

Temporal aggregation: one map per interval of a chosen granularity, with a statistic such as the mean

Core surface: the per-cell minimum over the series; near bare ground on a field

Envelope surface: the per-cell maximum over the series; the tallest the surface ever was

Cumulative viewshed: the number of observer points from which each cell is visible

• Midterm vocabulary, and the words the lab report has to use correctly
• The grading asks for conclusions in these terms: intersection, aggregation, core and envelope, distorted survey

Wrap-up

Assignment 5B: Analysis of multitemporal UAS data

• Recap: register once, then every analysis is a t.* tool on the dataset
• Core and envelope bracket everything the surface did; a core below bare ground points at a bad survey
• Map algebra on the series turns DSMs into the quantity you care about: crop height, then biomass
• Next: the guest lecture and the midterm, then Topic 6 on image classification and change detection

References

Gebbert, Sören, and Edzer Pebesma. 2014. “A Temporal GIS for Field Based Environmental Modeling.” Environmental Modelling & Software 53: 1–12. https://doi.org/10.1016/j.envsoft.2013.11.001.

Mitasova, Helena, Eric Hardin, Margery Overton, and Russell S. Harmon. 2009. “New Spatial Measures of Terrain Dynamics Derived from Time Series of Lidar Data.” 2009 17th International Conference on Geoinformatics, 1–6. https://doi.org/10.1109/GEOINFORMATICS.2009.5293539.