32
Near-real time monitoring of deforestation using a neural network and MODIS data: the PARASID approach Andy Jarvis, Louis Reymondin, Jerry Touval CIAT and TNC

Andy Jarvis and Louis Reymondin - PARASID Near Real Time Monitoring Of Deforestation Using A Neural Aug 2009

  • Upload
    ciat

  • View
    627

  • Download
    2

Embed Size (px)

Citation preview

Page 1: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Near-real time monitoring of deforestation using a neural

network and MODIS data: the PARASID approach

Andy Jarvis, Louis Reymondin, Jerry Touval

CIAT and TNC

Page 2: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Contents

• The approach• The implementation• Some examples• Comparison with DETER• Comparison with FORMA• Plans and timelines

Page 3: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Objectives of PARASID

HUman Impact Monitoring And Natural Ecosystems

• Provide near-real time monitoring of habitat change (<3 month turn-around)

• Continental – global coverage (forests AND non-forests)

• Regularity in updates

Page 4: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

The Approach

The change in greenness of a given pixel is a function of:

• Climate• Site (vegetation, soil, geology)• Human impact

Page 5: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Machine learning

We therefore try to learn how each pixel (site) responds to climate, and any anomoly corresponds to human impact

Machine learning (or neural-network), is a bio-inspired technology which emulates the basic mechanism of a brain.

It allows – To find a pattern in noisy dataset– To apply these patterns to new dataset

Page 6: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

4500

5000

5500

6000

6500

7000

7500

8000

8500

9000

1 2 3 4 5 6 7 8 9Time

NDVI

Measurments

Predictions

Interval max

Interval min

4500

5000

5500

6000

6500

7000

7500

8000

8500

9000

1 2 3 4 5 6 7 8 9Time

NDVI

Measurments

Predictions

Interval max

Interval min

NDVI Evolution and novelty detection

Novelty/Anomoly

Page 7: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

What is “machine learning” ?

• Machine learning (or neural-network), is a bio-inspired technology which emulates the basic mechanism of a brain.

• It allows – To find a pattern in noisy dataset– To apply these patterns to new dataset

Page 8: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Methodology

NDVIt

Precipitation (t)

Temperature(t)

w0

w1

w2

NDVI(t-1)

NDVI(t-2)

NDVI(t-n)

wp1

wp2

wp3

wo1

wo2

wo3

As required by the ARD algorithm, each input and the hidden output is a weights

class with its own α α0

αc

INPUTS: Past NDVI (MODIS 3b42) Previous rainfall (TRMM) Temperature (WorldClim)

OUTPUT: 16 day predicted NDVI

Page 9: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Methodology – Bayesian NN

• To detect novelties, Bayesian Neural Networks provide us two indicators– The predicted value– The probability repartition of where the value should

be

• The first one allows us to detect abnormal measurements

• The second one allows us to say how sure we are a measurement is abnormal.

Page 10: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

The Processing

• For South America alone, first calculations approximated 10 years of processing for the NN to learn:– A map of 30720 by 37440 pixels

1,150,156,800 vectors 23 vectors per year 26,453,606,400 NDVI values to manage per year 9.5 years of data 251,309,260,800 individual data points

• Through various processes, optimizations and hardware acquisitions reduced time to 3 months for NN learning

• Detection takes 2-3 days

Page 11: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

The Bottom-Line

• 250m resolution• Latin American coverage (currently)• 3 week turnaround from data being made

available (4 week delay in MODIS going to NASA ftp) (3+4 = 7 weeks)

• Report every 16 days• Measurement of scale of habitat change

(0-1) and probability of event

Page 12: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

An Example

• Two animations for Mato Grosso

Page 13: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009
Page 14: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID vs. DETERValidation tests

Page 15: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Validation area

• Location – Brazil– Bottom left corner

• 09°40’00’’ S • 60°00’00’’ O

• Size– 160 * 120 [km2]– 307200 modis pixels

Red square : Validation area

Page 16: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Data source

• Deter dataset has been downloaded on the website of the Deter project : – http://www.obt.inpe.br/deter/indexdeter.php?id=6424

• The satellite images of 2004 and 2006 used to compare the two models are extracted from images – Mosaico LandSat 2004 (AMZ)– Mosaico LandSat 2006 (AMZ)

• Also provide by the Deter project website.

Page 17: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID true positives

2004

2006

Parasid model is, sometimes more sensitive, and detects events that Deter doesn’t detect.

Page 18: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID True Positives

It seems Parasid model detects quite small and isolate events which Deter doesn’t detect.

2006

2004

Page 19: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID False positives

On the other hand, Parasid is more sensitive to false positives. Here, around a river.

2004

2006

Page 20: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID False Positives

In this example, Parasid doesn’t detect as well as Deter the big new field (red circle) but,is more precise to detect the small fields on the top right corner (blue circle) .

2004

2006

Page 21: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Synthesis

• The general locations of detections are the same for both models, so we have in essence created a DETER that works more automatically and is applicable at the continental level

• Both PARASID and Deter models have problems with false positive events, but we can adjust for these using this and other validation data and a genetic algorithm.

• PARASID seems more sensitive for small and isolated events, but this sensitivity may also generate false positives. Again, this is fixable through calibration data.

Page 22: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID vs. FORMAValidation tests

Page 23: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Test area

• Brazil– Latitude 8°36'7.50"S– Longitude 51°28'30.00“W

NDVI 2000.02.18 NDVI 2004.01.01 NDVI 2009.01.01

Page 24: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Models’ output

PARASID detectionsFirst detection in 2004

FORMA probabilitiesFirst detection in 2000

Page 25: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

PARASID – FORMA Red pixels show where FORMA’s probabilities are higher than the PARASID onesWhite pixels show where PARASID’s probabilities are higher than the FORMA’s ones

• This map shows the models’ differencies– The two models match

Comparison

Page 26: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Detailed comparison

NDVI 2000.02.18 NDVI 2004.01.01PARASID - FORMA

• The marked areas show that most of the differences are due to the changes which happened between 2000 and 2004.

• Parasid can’t detect these changes as it started detecting in 2004.

Page 27: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Detailed comparison

Top FORMA Bottom PARASIDImages from google earth

PARASID - FORMA

Maybe due to the rescaled pixel size from 250 [m] to 500 [m], FORMA model doesn’t fit perfectly some fields (the red bound around the fields on the comparison map).

Page 28: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Detailed comparison

Top FORMA Bottom PARASIDImages from google earth

PARASID - FORMA

Parasid is a bit more sensitive.

Page 29: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Detailed comparison

Clear change in 2006Softer change in 2008

Maybe vegetation degradation

The pixel plotted is shown in red on the map

.

Page 30: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Detailed comparison

• The two graphs in the previous slide show that the changes detected by PARASID and not by FORMA actually occurred (graph on the right).

• However these changes are more subtle than classic deforestation events (graph on the left).

• This could be due to vegetation degradation (e.g. selective logging) as the area where these changes occurred is surrounded by a high rural activity

• Also some benefits from 250m resolution• PARASID and FORMA complementary in provision of

monitoring

Page 31: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

Conclusions and next steps

• Near-real time global monitoring is possible• PARASID now functioning for Latin America• Next milestones:

– Fully functioning web interface January 2010– Continental validation and calibration (January 2010)– Global extent (2011)– Additional models to identify type of change (drivers)

(2011)

Page 32: Andy  Jarvis and Louis Reymondin - PARASID  Near Real Time Monitoring Of Deforestation Using A Neural  Aug 2009

More info…

[email protected]