sentinel-2 for agriculture

Theia Workshop for Sentinel-2 L2A MAJA products – 13 & 14 June 2018

Sentinel-2 for AgricultureS2 OPERATIONAL EXPLOITATION FOR SUPPORTING NATIONAL TO REGIONAL

AGRICULTURE MONITORING AROUND THE WORLD AT PARCEL LEVEL

Bontemps S., Defourny P., Bellemans N., Cara C., Dedieu G., Guzzonato E., Hagolle O., Inglada J., Nicola L., Rabaute T.,

Savinaud M., Udroiu C., Valero S., Koetz B.& 12 benchmarking & demonstration sites partners



• Major international attention about food supply (insecurity, price volatility)

• International land grabbing challenging national food security

Agriculture monitoring as a hot topic

S2 as an opportunity for agriculture

Need for better agricultural monitoring capabilities

EO observation can help

Agriculture monitoring as a hot topic


In the big data era providing dense and systematic time series and cloud computing facilities, the key question is :

can we exploit in near real time the EO data flows at parcel level over large areas dealing with the cropping systems diversity ?

S2 for operational agriculture

monitoring at 10 m resolution ?


Sentinel-2 for Agriculture systemUser-driven approach to define the requirements

Consortium Champion UsersSupport &data provider

Survey filed up by 42 institutions

1st User Consultation organized by ESA in 2012

2nd User Consultation through surveys in 2014

1st Sen2-Agri Users Workshop – FAO May 2014

2nd Sen2-Agri Users Workshop – EU Nov. 2015

3rd Sen2-Agri Users Workshop – FAO Jun. 2017


EARLY AREA INDICATOR

EARLY AREA INDICATOR

Binary map identifying annually cultivated land at 10m updated every month

Crop type map at 10 m for the main regional

crops including irrigated/rainfed

discrimination

Vegetation status map at 10 m delivered every

week (NDVI, LAI, pheno index)

Monthly cloud free surface reflectance composite at 10-20 m

Top priority : automatic delivery of 4 agri.

products along the season from S2 & L8 in line with the GEOGLAM core products


Benchmarking to select the best

algorithms for each product

4 publications

network is instrumental to validate over various sites

12 sites globally distributed to

select methods for cropland

and crop type mapping

To develop the open source

Sen2-Agri system to deliver

agriculture products at national

scale(Bontemps et al., 2015)

FRANCE MOROCCOMARICOPA (USA) ARGENTINASOUTH AFRICA

http://www.jecam.org/


Monthly cloud-free

composite

April, France

Weighting average approach

LAI products

• LAI retrieval by machine learning to build a non-linear regression model

• Reflectance simuated using the ProSail model• Mono-date LAI• Multi-date LAI

• Weekly product by weighted average usingthe n last LAI value

• At the end of the season, by fitting a phenological model on the full time series

Weighted Average

Synthesis Processor

(WASP)


Dynamic cropland

maskCrop type map

L2A time series

In situ data

Featuresextraction

Randomforest

Cropland / crop type

map

Reference LC map

Trimming(samplescleaning)

http://maps.elie.ucl.ac.be/CCI/viewer/index.php

CCI global LC map

2015 - 300 m - 22 classes

Matton et al., RS2015

Valero S. et al., RS2016

Inglada et al., RS2015


Sen2-Agri open source system

Sen2Agri Ochestrator

S2 L1C Time Series

L8 L1T Time Series

L2A Time Series

AtmosphericCorrections

(MACCS)

Temporal Synthesis Processor

Temporal Synthesis ProcessorNode #1

Biophysical Indicators Processor

Biophysical Indicators ProcessorNode #2

Cropland Mask Processor

Cropland Mask ProcessorNode #3

Crop Type Map Processor

Crop Type Map ProcessorNode #4

In-situInformation

Surface Reflectance Composite

Vegetation Status

Indicators

Dynamic Annual

Cropland Mask

Crop Type Map & Extent Area

Automated Download

Scheduled

Scheduled

Scheduled

Scheduled

Scheduled

Triggered

NRT or off-line production, locally or in the cloud,

automated or manual production


Sen2-Agri system : main parameters settings

Area of Interest Shapefile to be uploaded

Monitoring period Start and end dates to be defined

S2 or S2 + L8 To be selected

Other parameters …

Start of the season

Monitoring periodBefore the monitoring period

System initialization

End of the season…

Sen2-Agri system : field campaigns

Sampling design Stratification and sampling

Field visit In situ data collection – early survey

In situ data collection – mid-season survey

Data upload Field data quality control and formating

Mali stratification (PIRT)

Simple parametrization

Field data collection planning


Monitoring period

System operation for crop growth monitoring

LAI product

EoSSoS

Automatic EO data downloadfrom EO data providers

SoS

Before the monitoring period

t0

t2

t5 t10 t15

t18

t0t2

t10

t15

t18

System initialization


17/07/2016

26/08/2016

05/09/2016

05/10/2016

25/10/2016

15/11/2016

0

1

2

3

4

Koutiala

2016-2017 LAI time series at 10m

(Mali)


2016-2017 LAI time series at 10m

(South Africa)

16/11/2016

26/11/2016

06/12/2016

26/12/2016

15/01/2017

14/04/2017

0

1

2

3

4

5

Potchestroom (South-Africa)

27/10/2016

Leaf Area

Index


Cloud-free temporal synthesis

South of France, S2 + L8


Juillet 2016

Août 2016

Septembre 2016

Octobre 2016

Cape Town

Western Cape Province monitored by Sentinel 2 in 2016

June July August September October November

Winter grain production region (South Africa)


Sen2-Agri 10 m cropland map for Ukraine

September 2016

Overall accuracy : 98 %

F-score cropland : 97 %

Non-Cropland

Cropland

November 2016

Overall accuracy : 98 %

F-score cropland : 99 %


Sen2-Agri 10 m main crop types map for Ukraine (July 2016)

Overall accuracy : 82,7 %

NRT crop specific monitoring from the mid-season:

LAI evolution for all winter wheat fields in 2016


Comparison of TOA, Sen2COR, MACCS for crop type classification

Experiment on L2A algorithms

by Ukrainian team (SRI)

Satellite data:

11 Sentinel-2 scenes

Ground data:

563 ground samples

(train and test sets)

Training set

Test set

TOA (OA=80.7%) Sen2cor (OA=80.6%) MACCS (OA=82.7%)


Monthly cropland product performance

along the season

F-Score Cropland F-Score Non-Cropland Overall Accuracy


Overall accuracy: 94 %

F-score cropland: 80 %

System validation:

2016 Cropland mask at 10m resolution for Mali

from Sentinel-2 and Landsat 8


Fast learning curve to improve performances

Mali – work continued with our partners


Sen2-Agri 10 m main crop types map for South AfricaSummer grain production area – Northern provinces

Maize

Soyabeans

Sunflower

Fodder crops

Other crops

Free State

Main Crops F1-Score

Maize 94%

Soyabean 83%

Sunflower 82 %

Fodder crops 90%

Overall Accuracy 90,1 %EO data: 6 months of S2 and

L8 acquisitions(Oct 2016 – March 2017)

In situ data: ~12000 samples(~451000 ha)

Sen2-Agri Online Training for USDA – 19 April 2018 22/30


Open-source, available for download,

demonstrated at full scale, user community

Any EO users to produce agriculture

information in their area of interest: national service for agriculture monitoring

national statistical offices

international agencies

private companies

research organizations

400 downloads

20 operational installations

http://forum.esa-sen2agri.org/

Very active forum with

60 users

1.1 K posts

150 topics

Webinar every 2 months

-> Next one: 19 July

Online trainings

http://forum.esa-sen2agri.org/


Thank you for your attention

… and come and play with us!http://www.esa-sen2agri.org/


Sen2-Agri system: S2 exploitation for local to national

operational agriculture monitoring at 10 m

Decametric time series availability allow to proceed to full scale experimenton annual basis – Sentinel-2b much needed and Sentinel-1 integration wouldclearly enhance the timeliness and the robustness

Nationwide Sen2Agri NRT operation requires good processing but more critically in situ data collected and quality controlled in a timely manner(logistics matter a lot - flooding in Sudan, unrest in Madagascar, etc.)

Very fast learning curve thanks to this challenging globally distributeddemonstration allowing to further improve the Sen2Agri system optimizingoperational performances (timeliness, accuracies in challenging landscapes)

Sen2Agri system is a turn-key system able to evolve in close partnerships with national stakeholders to insure outcome relevance and ownership -multi-year experience much needed to fully exploit Sentinel-2 capabilities !


Sen2-Agri is free and open source

Based on open source existing software

Under GNU-GPL License

Open source software based on Orfeo ToolBox

Cluster-ready architecture for distributed processing

Integration with Sentinel-2 ToolBox

Operational system required : CentOS7 (GNU/LINUX)


• Optional to generate the cropland mask

• Mandatory to generate the croptype map

Field campaigns for cropland and croptype can be combined.

The quality of the in situ data set is the most important driver of the quality of the output maps

With the proper format, one single dataset can be given to the system.

2 different objectives:

In situ data for calibration/training of the machine learning algorithms: sampling to cover the diversity of the existing classes accross the site. It should represent the range of possible land cover types and crop types.

In situ data for validation: to estimate the model performance to recognize the different classes. To be consider as a proper map validation, the validation dataset should be collected using a statistically-sound sampling.

In situ data collection

By-default split of the dataset:75% of the samples to support the training25% of the samples to test the performance of the classifier


• Shapefile to be provided with specific attribute table format [cfr Software User Manual]

• Stratification should be applied in case of large area mapping

• Guidelines for field campaign method are provided by the JECAM network:http://www.jecam.org/JECAM_Guidelines_for_Field_Data_Collection_v1_0.pdf

Dedicated webinar on in situ data collection methods is planned on July 19th

In situ data preparation

http://www.jecam.org/JECAM_Guidelines_for_Field_Data_Collection_v1_0.pdf


If you intend to use the Sen2-Agri system on a local site (~90 000km²), the minimum system requirements to install and run this system are as follows:

• CPU: 8 Cores

• RAM: 32GB

• HDD Storage: 4 TB

• SSD Storage: 250 GB (optionnal – for temporary files)

If you intend to use the Sen2-Agri system on a national site (~500 000km²), the minimum system requirements to install and run this system are as follows:

• CPU: 16 to 32 Cores

• RAM: 64GB to 128 GB

• HDD Storage: 10 TB is enough even if 15TB is recommended

• SSD Storage: 1 TB (optionnal – for temporary files)

System requirements


Amount of downloaded product (L1C Sentinel-2 and L1T Landsat 8)

For the system initialization (multi-temporal atmospheric correction)

S2: 63 products, 226 GB

L8: 327 products, 621 GB

For the growing season until today

S2: 427 products – 1.341TB

L8: 623 products – 1.2TB

Amount of L2A generated

For the system initialization (multi-temporal atmospheric correction)

S2: 63 products, 756 GB (2 months)

L8: 146 products, 205 GB (3 months)

For the growing season until today

S2: 427 products – 5.33 TB

L8 : 554 products – 775 GB

High level figures :

Ukraine national demonstration


Processing time for each Sen2Agri product

• For Atmospheric correction: Around 30 minutes per tile

• For Cloud free composite: 8 days for a product with S2 and L8

• For LAI: Approx. 3 minutes/ S2 tile

• For L4A: About 3 days for the raw mask, about 5 days and a half for segmentation

• For L4B: About 8 days

National hardware server : 2 x Intel(R) Xeon(R) CPU E5-2650 v3 @ 2.30GHz (20 cores total, 40

threads), 128 GB RAM (26 TB disk space)

System performances over national sites


Madagascar site – mid-season cropland and

crop type products thanks to timely in situ data


201

7


Madagascar site – mid-season cropland and

crop type products thanks to timely in situ data


November 201650 days window

5/10/2016-25/11/2016

First cloud free composite series over Mali

at 10m resolution from Sentinel-2 and Landsat 8


Cropland product for local site in Northern China

without in situ data (trained from global CCI LC map)

Cropland mask, obtained without in-situ data but using the ESA CCI Land Cover map to extract training dataset(http://maps.elie.ucl.ac.be/CCI/viewer/)

Cropland

Non cropland


Monthly cloud-free composite

Benchmarking conclusions

• Weighted average approach

• Compositing period can vary between 30 to 50 days window

• Implements a directional correction for seamless composites

• Recurrent implementation : L3A product updated with each new L2A product

To limit the data volume to keep on-line in Sen2Agri system

April, France


Dynamic cropland mask2 chains implemented to deal with presence/absence of in-situ data

Benchmarking conclusions• Surface reflectance preparation

• Linear interpolation of the refl data for the in situ data version• Gap filling using Whittaker smoothing if absence of in situ data

• RF supervised algorithm on a set temporal feature• Trimming to clean the reference map• A posteriori smoothing based on a per-object approach

(Matton et al., RS2015)

(Valero S. et al., RS2016)


Crop type map

Benchmarking conclusions• Surface reflectance preparation

• Linear interpolation of the refl data for the in situ data version• Based on the crop mask previously generated• Random forest classifier• Classifier applied on temporal features : surface reflectance, NDVI,

NDWI, brightness

(Inglada et al., RS2015)

sentinel-2 for agriculture

Documents