Assessing the Limitations and Capabilities of Lidar and Landsat to Estimate Aboveground Vegetation

Biomass and Cover in Semi-arid Rangeland Ecosystem Using a Machine Learning Algorithm

Shital DhakalMS in Hydrologic Sciences

Semi-arid ecosystemArid & Semi arid cover 1/3rd of Earth’s land surface

These ecosystems are fragile due to water limitations

Important roles in ecology, hydrology and carbon cycling

(Pho

to b

y L.

Eng

ledo

w)

(Pho

to b

y S

. Har

degr

ee)

Semi-arid in NW United StatesHistorically dominated by sagebrush steppe

One of the most important plants on western rangelands

Ecological, hydrological and carbon cycle importance

An imperiled ecosystem!

Reasons are increased fire frequency and replacement by exotic plants

Field based quantification Quantification is essential for modeling ecosystem, conservation of vegetation,

measuring fuel, understanding carbon flux etc.

Some in-situ measurement methods are available

E.g. Harvesting, Clip and Weigh, Visual estimation, Point intercept sampling

Destructive method (clipping- oven drying) are expensive and labor intensive

Point intercept sampling involves taking multiple field measurements

Big problem- Very small geographical extent!

Can remote sensing provide the answer?

Previous studies

Hypothesis Satellite spectral remote sensing and Airborne Lidar can independently provide

large scale accurate estimation of vegetation characteristics in semi-arid rangeland.

Research questions What lidar based metrics are best to estimate biomass and cover of semi-arid vegetation ?

Is point cloud processing of lidar data better than lidar based raster data?

How does spectral remote sensing compare with lidar based estimation?

What methods can be best utilized to produce biomass and cover map across a large scale?

Can the available remote sensing technologies be used to estimate characteristics of both herbs and shrubs in rangeland?

Outline of methodology - Study site

- Field Data Collection

- Lidar data & Lidar data processing

-Landsat & Landsat image processing

-Random Forest

- Results from Lidar

-Results from Landsat

-Lidar vs Landsat

Study site

Morley Nelson Snake River Birds of Prey NCA, Boise

Data collection

Field data distribution (n = 141)

0 10 20 30 40 50 60 70 80 90 1000

10

20

30

40

50

60

70

Percentage Cover

Freq

uenc

y of

fiel

d da

ta

100 200 300 400 500 600 100080

60

40

20

0

20

40

60

80

100 Shrub Herb

Biomass (g/m2)

Fequ

ency

of fi

eld

data

Lidar data

Leica ALS 60 Lidar Sensor

65,000 hectares in 2012 and 9,000 hectares in 2013

Discrete small footprint lidar data

Point density of ~ 8 points per sq. meter

Vertical accuracy of ~ 3 cm

Flown at 1,500 m , acquiring ≥ 148,000 laser pulses per second

Leica ALS 60 Lidar sensor

Lidar data processingLidar derived metrics

◦ Minimum Height ◦ Maximum Height◦ Height Range◦ Mean Height ◦ Median Absolute Deviation (MAD) from Median Height ◦ Mean Absolute Deviation (AAD) from Mean Height ◦ Height Variance◦ Height St. Deviation ◦ Height Skewness◦ Height Kurtosis ◦ Interquartile Range (IQR) of Height◦ Height Coefficient of Variation◦ Height Percentiles - 5th, 10th, 25th, 50th, 75th, 90th,& 95th ◦ Number of Lidar Returns ◦ Number of Lidar Vegetation Returns (nV) ◦ Number of Lidar Ground Returns (nG) ◦ Total Vegetation Density ◦ Vegetation Cover ◦ Percent of Vegetation in Height Range ◦ Canopy Relief Ratio ◦ Texture of Heights◦ Foliage Height Diverstiy (FHD)

Buffering

Height Filtering

Raster

Point Cloud

1 m

7 m

30 m

100 m

Landsat 8 OLI Images the entire Earth every 16 days

Spatial Resolution of 30 m

Pushbroom Sensor

Generates 16-bit images

Eight fold increase in signal-to-noise ratio

Landsat 8 OLI Images the entire Earth every 16 days

Spatial Resolution of 30 m

Pushbroom Sensor

Generates 16-bit images

Eight fold increase in signal-to-noise ratio

Data Preparation (n=97+44)

11 Apr 2013

30 Jun 20134 Oct 2013

Topographic variables

Slope

Elevation

Aspect

More on vegetation indicesSimple Ratio based VI-E.g. NDVI, SCI, VCI

Soil adjusted VI-E.g. SAVI, GSAVI, SATVI

Perpendicular VI-E.g. BI, GVI, WI

Why random forest (RF)? One of the most accurate machine learning algorithms available

Not affected by multi-collinearity between variables

Well suited for analyzing complex non-linear dataset

Well suited for broad data i.e many predictors but low sample size

Gives estimate of what variables are important

Lidar model Landsat model

Sample Size 46 141

No. of Predictor variables 35 81

No. of Target variables 3 5

Data set from NCA:

RF lidar raster analysis Area Resolution R2 RMSE Best Predictors

Total biomass

1 ha 1 m

0.74 141 Hstd, HAAD, H90, HSkew, Hvar, Htext

1 ha 7 m

0.70 152 Htext, FHDGT, H95, HAAD

1 ha 30 m

0.58 180 FHDGT, nV, HAAD, H5

1 ha 100 m

0.52 188 FHDGT, nV, H16, HAAD

Shrub biomass

1 ha 1 m

0.76 152 Hstd, HAAD , HCV, Hrange, FHDall

1 ha 7 m 0.67 143 Htext, FHDGT, HAAD

1 ha 30 m

0.50 176 FHDGT, HAAD, HCV

1 ha 100 m 0.4 184 Htext, H50, pG, nG

RF lidar raster analysis Area Resolution R2 RMSE Best Predictors

Total biomass

1 ha 1 m

0.74 141 Hstd, HAAD, H90, HSkew, Hvar, Htext

1 ha 7 m

0.70 152 Htext, FHDGT, H95, HAAD

1 ha 30 m

0.58 180 FHDGT, nV, HAAD, H5

1 ha 100 m

0.52 188 FHDGT, nV, H16, HAAD

Shrub biomass

1 ha 1 m

0.76 152 Hstd, HAAD , HCV, Hrange, FHDall

1 ha 7 m 0.67 143 Htext, FHDGT, HAAD

1 ha 30 m

0.50 176 FHDGT, HAAD, HCV

1 ha 100 m 0.4 184 Htext, H50, pG, nG

RF lidar point cloud analysis Area Resolution R2 RMSE Best Predictors

Total biomass

1 ha 1 m

0.71 147 HMAD, HSkew, HIQR, HAAD, Hstd, Hkurt, H90, HCV

1 ha 7 m

0.71 148 Htext, HIQR

1 ha 30 m

0.70 151 HAAD, H95, HIQR, pH1,pG

1 ha 100 m

0.67 160 H90, H95, Htext, veg_density

Shrub biomass

1 ha 1 m

0.73 129 HIQR, Hstd , HMAD, HCV

1 ha 7 m 0.72 132 Htext, H90, HIQR, HCV

1 ha 30 m

0.65 146 H90, HIQR, Htext, pH1

1 ha 100 m 0.64 151 H95, Htext, pH1, GIQR,FHDGT

RF point cloud analysis Area Resolution R2 RMSE Best Predictors

Total biomass

1 ha 1 m

0.71 147 HMAD, HSkew, HIQR, HAAD, Hstd, Hkurt, H90, HCV

1 ha 7 m

0.71 148 Htext, HIQR

1 ha 30 m

0.70 151 HAAD, H95, HIQR, pH1,pG

1 ha 100 m

0.67 160 H90, H95, Htext, veg_density

Shrub biomass

1 ha 1 m

0.73 129 HIQR, Hstd , HMAD, HCV

1 ha 7 m 0.72 132 Htext, H90, HIQR, HCV

1 ha 30 m

0.65 146 H90, HIQR, Htext, pH1

1 ha 100 m 0.64 151 H95, Htext, pH1, GIQR,FHDGT

RF lidar analysis (Herb Biomass) Area Source Resolution R2 RMSE Best Predictors

Herb biomass

1 ha Raster 1m

0.2 6.86 HSkew, Htext

1 ha Point Cloud

1m

0.19 7.54 HCV, Htext, HSkew

RF Landsat analysis  Calibration (n=97) Validation (n=44)

  R2 RMSE Variables R2 RMSEShrub Cover 0.63 7 June30 SATVI, June30 GVI,

Oct4 SAVI, Oct4 MSAVI,June30 NBR

 

0.44 8

Herbaceous Cover 0.69 13 Oct4 MIRI, June30 MSI,Oct4 Green,June30 SATVI

 

0.63 16

Total Biomass 0.54 147 June30 SATVI, June30 GVI,Oct4 NIR, Oct4 MSAVI,Oct4 Red, Oct4 SAVI

 

0.37 158

Shrub Biomass 0.6 126 June30 SATVI, June30 GVI, Oct4 MSAVI, June30 NDWI,Oct4 GSAVI

 

0.53 128

Herb Biomass 0.49 65 Apr11 NIR, June30 MSI,June30 NIR, Oct4 NIR

0.3 64

RF Landsat analysis  Calibration (n=97) Validation (n=44)

  R2 RMSE Variables R2 RMSEShrub Cover 0.63 7 June30 SATVI, June30 GVI,

Oct4 SAVI, Oct4 MSAVI,June30 NBR

 

0.44 8

Herbaceous Cover 0.69 13 Oct4 MIRI, June30 MSI,Oct4 Green,June30 SATVI

 

0.63 16

Total Biomass 0.54 147 June30 SATVI, June30 GVI,Oct4 NIR, Oct4 MSAVI,Oct4 Red, Oct4 SAVI

 

0.37 158

Shrub Biomass 0.6 126 June30 SATVI, June30 GVI, Oct4 MSAVI, June30 NDWI,Oct4 GSAVI

 

0.53 128

Herb Biomass 0.49 65 Apr11 NIR, June30 MSI,June30 NIR, Oct4 NIR

0.3 64

RF Landsat analysis   Calibration (n=97) Validation (n=44)

  R2 RMSE Variables R2 RMSEShrub Cover 0.63 7 June30 SATVI, June30 GVI,

Oct4 SAVI, Oct4 MSAVI,June30 NBR

 

0.44 8

Herbaceous Cover 0.69 13 Oct4 MIRI, June30 MSI,Oct4 Green,June30 SATVI

 

0.63 16

Total Biomass 0.54 147 June30 SATVI, June30 GVI,Oct4 NIR, Oct4 MSAVI,Oct4 Red, Oct4 SAVI

 

0.37 158

Shrub Biomass 0.6 126 June30 SATVI, June30 GVI, Oct4 MSAVI, June30 NDWI,Oct4 GSAVI

 

0.53 128

Herb Biomass 0.49 65 Apr11 NIR, June30 MSI,June30 NIR, Oct4 NIR

0.3 64

Nearest neighbor imputation Replacing missing data with substituted values In comparison Interpolation predicts values at un-sampled locations Biomass is produced as weighted averages of selected variables The variables are selected by Random Forest The reference data should cover the entire phenomenon of interest We used yaimpute package in R for Imputation

Lidar biomass maps

Landsat biomass maps

Landsat biomass maps

100

200

300

400

500

600

600+

60 40 20 0 20 40 60 80

Percent distribution

Biom

ass (

g/m

2)

Landsat cover maps

10

20

30

40

50

60

70

80

90

100

20 10 0 10 20 30 40 50 60 70 80

Percent distribution

Perc

ent C

over

Landsat vs Lidar  Landsat 8 OLI Lidar

  R2 RMSE Variables R2 RMSE VariablesShrub Cover 0.75 6.5 June30 SATVI,

Oct4 GSAVI,Apr11 DVI

 

0.74 6.7 Hrange, FHDAll

Herbaceous Cover 0.6 12.5 Apr11 DVI,June30 MSI, June30 VCI,June30 SATVI, Oct4 BI

 

0.21 17.5 Hrange, HIQR

Total Biomass 0.57 177 Oct4 BI, Oct4 NIR, Elevation, Oct4 SWIR,Oct4 MSAVI

 

0.68 156 FHDAll, Hstd,AAD, Hrange, HSkew

Shrub Biomass 0.61 151 June30 DVI, Oct4 BI,Elevation, Apr11 DVI

 

0.75 126 Hstd, Hrange,FHDAll, HCV

Herb Biomass 0.57 57 June30 GSAVI,June30 MSAVI,June30 SWIR, Oct4 GVI

0.12 83 H10, HSkew, CRR,AAD

Summary I 1) The vegetation cover and biomass of shrubs in large scale was successfully modeled using

multispectral imagery (Landsat 8) and airborne lidar.

2) We found that the best model to describe vegetation cover fractions included vegetation indices

calculated from multiple acquisitions dates.

3) Lidar was found to estimate shrub biomass slightly better than Landsat.

4) Validation is necessary for reducing the bias

Summary II 4) Point cloud processing of lidar data significantly improves the estimation of biomass in

coarser scale compared to raster processing.

5) Lidar could not satisfactorily model the herbaceous biomass in the field site (R2 < 2).

6) As per our imputed map, the NCA contains ~ 345 metric ton of herbaceous biomass

and ~ 313 metric ton of shrub biomass.

An application - fire events

0 1 2 3 4 5 6 7 8 90

40

80

120

160

200 Herb

Shrub

Fire frequency

Aver

age

biom

ass (

g/m

2)

0 1 2 3 4 5 6 7 8 90

20

40

60

80 Herb

Shrub

Fire frequency

Aver

age

cove

r (%

)

0 1 2 3 4 5 6 7 8 9 100

50

100

150

200

250

300

Fire frequency

Aver

age

biom

ass (

g/m

2)

0 1 2 3 4 5 6 7 8 9 100

102030405060708090

100

Herb

Fire frequency

Aver

age

cove

r (%

)

With imputed pixels

With in-situ field plots

Conclusion We suggest to use lidar for biomass estimation and Landsat for herbaceous cover estimation Remote sensing can extended field based methods across larger scaleThe application of the methodology is wide. From land managers to ecologistSynergetic use of remote sensing data in future can produce better results.Can be repeated in other part of world with some minor tweak

Acknowledgement Committee chair- Dr. Nancy F. Glenn

Subject Committee – Dr. Alejandro N. Flores, Dr. Douglas J. Shinneman, Dr. Aihua Li

Dr. Rupesh Shrestha, Lucas Spaete

Peter Olsoy, Kyle Anderson, Hamid Dashti, Reggie, Ann Marie, Andrew, Katie

and Ginikanda Yapa Mudiyanselage Nayani Thanuja Ilangakoon

Nepalese friends and everyone in Geosciences!

Questions?

Regression Analysis

0 200 400 600 800 1000 12000

100

200

300

400

500

600

700

800

900f(x) = 0.784685456219575 x + 75.5139754762692R² = 0.785026053341471

Predicted total AGB (g/m2)

Obs

erve

d AG

B (g

/m2)

Field data distribution (n = 46) Herb Cover (%) Shrub Cover (%)

Herb Biomass (g/m2) Shrub Biomass (g/m2)

Minimum 0 0 2 0

Maximum 100 87 1207 3301

Mean ± SE 39 ± 1.47 12 ± 0.85 144 ± 7 414 ± 20

100 200 300 400 500 600 600+25

15

5

5

15

25 Herb Shrub

biomass (g/m2)

Freq

uenc

y of

fiel

d da

ta

RF Analysis (70 x 70 Plot) Area Resolution R2 RMSE Best Predictors

Total biomass

70m x 70m 1 m

00.68 156 FHDall, Hstd, HAAD, Hrange, HSkew

Shrub biomass

70m x 70m 1 m

00.75 126 Hstd, Hrange, FHDall, HCV

extras