Lin Yan & David Roy

Geospatial Sciences Center of Excellence

South Dakota State University

February 13-15th 2018

Advances in Emerging Technologies and Methods in Earth

Observation for Agricultural Monitoring Workshop

Beltsville, Maryland

Field size estimation, past and future opportunities

Why Study Field Sizes ?The size of agricultural fields is:

• a fundamental description of rural

landscapes

• of biophysical, ecological and

economic importance

• indicative of the degree of agricultural

capital investment, mechanization,

and labor intensity

• through time is indicative of

agricultural changes

Farmers ripping up fence line and clearing tree lot to increase field size

(5th September 2015, one mile north of Brookings, SD)

Photograph by David Roy

Importantly Landsat has both suitable resolution & length of record

to capture field size changesin many parts of the world

Landsat 30m observations since 1982

What would this histogram

look like for:

• all CONUS ?

• every decade ?

Conterminous United States (CONUS) Motivation

Histogram from 109,000 agricultural

fields (> 1 acre) digitized from Landsat

data acquired in 1977 – 1980 over

parts of the Mid-West and Canada

M. C. Ferguson and C. D. Badhwar, 1986. Field size distributions

for selected agricultural crops in the United States and Canada,

Remote Sensing of Environment, 19 (1), 25-45.

Field Extraction Methodology

Computational methodology designed to be

• fully automated

• no training data

• no human interactions

• needed for CONUS 30m Landsat application

Example WELD Weekly Products

Week 27: July 8 - 14 2008

10 Years of Alaska and CONUS

Landsat 7 ETM+ 30m products

http://weld.cr.usgs.gov

Use Landsat time series

gridded calibrated 30m Landsat reflectance

weekly, monthly, seasonal and annual products

Example WELD Weekly Products

Week 28: July 8 - 14 2008

10 Years of Alaska and CONUS

Landsat 7 ETM+ 30m products

http://weld.cr.usgs.gov

Use Landsat time series

Example WELD Weekly Products

Week 29: July 15 - 21 2008

10 Years of Alaska and CONUS

Landsat 7 ETM+ 30m products

http://weld.cr.usgs.gov

Use Landsat time series

USDA National Agricultural Statistics ServiceCropland Data Layer (CDL)

based on supervised classification of many satellite data, lots of training data, and interactive refinements

Pixel-based products unable to extract separated and coherent fields

Big-picture algorithm processing flow

52 weeks of WELD 2010

Landsat 5 & 7 weekly images

USDA NASS 2010

Cropland Data Layer (CDL)

edge intensity map

Binary crop mask

Extracted fields

Computer vision approach

Detailed processing flow – computer vision approach

Yan, L. and Roy, D.P. (2016). Conterminous United States crop field size quantification from

multi-temporal Landsat data, Remote Sensing of Environment, 172, 67-86

Field Extraction Results

Texas

5000 x 5000 30m pixels

WELD Tile h13v12 Northern High Plains, Texas

Automatically extracted field objects

Texas

5000 x 5000 30m pixels

1200 x 900 30m pixels

1200 x 900 30m pixels

California

5000 x 5000 30m pixels

WELD Tile h05v13Imperial Valley, CA

Automatically extracted field objects

California

5000 x 5000 30m pixels

2200 x 1650 30m pixels

1400 x 1050 30m pixels

1400 x 1050 30m pixels

Field size

= (30 × 30) 𝑚2σ𝑛𝑢𝑚𝑏𝑒𝑟 𝑝𝑖𝑥𝑒𝑙𝑠

2010 CONUS crop field size map

km2

4,182,777 crop fields

extracted

(mean field size in 7.5 x 7.5 km grid cells)

derived from all 13,666 sunlit Landsat 5 and 7 scenes available in the U.S. Landsat archive for December 2009 to November 2010

Validation

Validation Sites

99,177 km2

Harvested area

• 48 sites distributed in the top 16 U.S. states by harvested area

• each site ~ 7.5 x 7.5 km

• >5,800 reference fields manually selected from Landsat 5 and Google-

Earth images over the 48 sites for comparison with the automatically

extracted fields

pixel-based accuracy metrics

Extracted fields

Site Ref. # Extr. # one to one

matching

ratio

over

-split

under-

split

Ref. mean

size

Extr. mean

size

mean

size diff

CA3 161 149 139 86.3% 1 8 318.7 315.7 -0.9%

Validation example: a California site ~ 7.5 x 7.5 km

Site fields PA fields UA OA

total Ref.

pixels

total Extr.

pixels

total pixels

diff

IA 1 90.4% 98.6% 93.5% 51315 47040 -8.3%

object-based accuracy metrics

USDA NASS CDL, 2010Extracted fields

& reference field polygons

Extracted fields

& reference field polygons

(one-to-one matched,over-split,

under-split,missed)

Landsat 5 image, acquired

7/13/2010

Validation results over 48 validation sites• 81.4% of the >5,800 reference fields correctly extracted• mean of <2% mean-field-size difference with <5% standard deviation

Extracted fields with reference field polygons

(one-to-one matched, over-split, under-split, missed)

Landsat 5 RGB image, acquired 8/23/2010

Google-

Earth image,

acquired

9/28/2010

Site Ref. # Extr. #

one to

one

matching

ratio

over

-split

under-

split

Ref.

mean

size

Extr.

mean

size

mean

size diff

MI 1 134 134 96 71.6% 14 13 464.9 447.5 -3.7%

Validation example: a Missouri site ~ 7.5 x 7.5 km

object-based accuracy metrics

CONUS 2010 crop field size histogram

4,182,777 fields extracted

Field size (km2)

Num

ber

of

field

s x

10

5

Fie

ld a

rea p

erc

en

tage (

%)

Field size (km2)

1 × 1/2 mile2

1/2× 1/2 mile2

1/2× 1/4 mile2

1/4 × 1/4 mile2

field area percentage =

Σ (field area in histogram bin)

Σ (CONUS field area)

CONUS 2010 crop field size histogram

4,182,777 fields extracted

California 2010 crop field size histogram

116,888 fields extracted

Area (km2)

Num

ber

of

fie

lds

Google-Earth image. ~5.5 x 5 km

subset in California near Corcoran

1 × 1/2 mile2

1/2× 1/2 mile2

1/2× 1/4 mile2

1/4 × 1/4 mile2

Fie

ld a

rea p

erc

en

tage (

%)

Field size (km2)

Iowa 2010 crop field size histogram

308,917 fields extracted

Num

ber

of

fie

lds

Area (km2)

Google-Earth image. ~5.5 x 5 km

subset in Iowa near Eagle Grove.

1 × 1/2 mile2

1/2× 1/2 mile2

1/2× 1/4 mile2

Fie

ld a

rea p

erc

en

tage (

%)

Field size (km2)

1/4 × 1/4 mile2

2010 CONUS CDL majority crop map(with ≥ 10% CDL crop pixels in 7.5 x 7.5 km grid cells)

soybeanscorn

alfalfa

wheat (winter, spring and durum wheat)cottonother crops

2010 CONUS crop field size map

km2

4,182,777 crop fields

extracted

(mean field size in 7.5 x 7.5 km grid cells)

derived from all 13,666 sunlit Landsat 5 and 7 scenes available in the U.S. Landsat archive for December 2009 to November 2010

CONUS 2010 field size histograms for the major crops

Fie

ld a

rea p

erc

en

tage (

%)

Field size (km2)

cornsoybeanalfalfawheatcotton

Largest Extracted Field

Google-Earth image

(acquired on 8/18/2010)

2010 CDL

(CDL classified as

cotton in the annual

2008 to 2014 product)

3,200 acres

(5 square miles!)

Gaines, Texas

Future Work

(but needs funding …)

Fritz et al. (2015). Mapping global cropland and field size, Global Change Biology

• field size categories from geo-wiki information • qualitative field size information only and unknown quality

Global Field Extraction

HDF format products at: http://globalweld.cr.usgs.gov/collections

GeoTiff format products at: http://globalweld.cr.usgs.gov

Native resolution visualizations at: http://go.nasa.gov/2kLcKto

GLOBAL WELD 3 years (2009-2011) of monthly & annual Landsat 5 & 7

composites atmospherically corrected Nadir BRDF-Adjusted Reflectance

(NBAR)

Landsat 8

400 × 400

30 m pixels

August 23 2016

California

Sentinel 2A

1200 × 1200

10 m pixels

August 23 2016

California

Li Z., Zhang H.K., Roy D.P., Yan L., Huang H., Li J., 2017, Landsat 15-m Panchromatic-Assisted

Downscaling (LPAD) of the 30-m reflective wavelength bands to Sentinel-2 20-m resolution,

Remote Sensing, 9(7), 755.

Landsat 8 LPAD

600 × 600

20 m pixels

August 23 2016

California

Landsat 8

400 × 400

30 m pixels

August 23 2016

California

Some fields too small to be discernable with Landsat or Sentinel-2 data

India – Punjab, 15 x 15km scene

Landsat 7 ETM+ 30m (10/28/2002) Quickbird-2 2.5m (10/07/2003)

Some fields too small to be discernable with Landsat or Sentinel-2 data

China – Jiangsu province, 15 x 15km scene

Landsat 5 TM 30m (03/23/2005) Quickbird-2 2.5m (04/07/2005)

Summary• Landsat time series provide sufficient information to detect crop fields in an

automated way using a computer vision based approach across the U.S.

• First-ever U.S. wall-to-wall satellite-based field extraction demonstrated (using WELD processed Landsat 5 and 7).

• Validation results over 48 validation sites

– 81.4% of the >5,800 reference fields correctly extracted

– mean of <2% mean-field-size difference with <5% standard deviation

• New moderate resolution satellite data will provide improved global agricultural monitoring where field sizes are small

– Landsat-8 30m data have better quantization and signal/noise characteristics than previous Landsats

– Sentinel-2 has Landsat-like bands at 10m & 20m

• We have been contacted by the National Geospatial-Intelligence Agency

(NGA) to investigate the approach on NGA commercial high-resolution

data under a Cooperative Research and Development Agreement

(CRADA).

AcknowledgementsResearch funded by NASA NNH09ZDA001N-LCLUC “Changing Field Sizes of the Conterminous United States, a Decennial Landsat Assessment”.

The USGS EROS are thanked for provision of the Landsat data and the USDA NASS are thanked for provision of the Crop Data Layer product.

References• Yan, L. and Roy, D.P. (2016). Conterminous United States crop field size

quantification from multi-temporal Landsat data, Remote Sensing of Environment, 172, 67-86.

• Yan, L. and Roy, D.P. (2014). Automated crop field extraction from multi-temporal Web Enabled Landsat Data, Remote Sensing of Environment, 144, 42-64.

• White, E. and Roy, D.P. (2015). A contemporary decennial examination of changing agricultural field sizes using Landsat time series data, Geo: Geography and Environment. DOI: 10.1002/geo2.4.