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

Contents The approach The implementation Some examples Comparison with other models Plans and timelines

The approach

The implementation

Some examples

Comparison with other models

Plans and timelines

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

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

The Approach The change in greenness of a given pixel is a function of: Climate Site (vegetation, soil, geology) Human impact

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

Climate

Site (vegetation, soil, geology)

Human impact

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

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

NDVI Evolution and novelty detection Novelty/Anomoly

NDVI Cleaning using HANTS Eliminate all short-term variations Uses NDVI quality information Iterative fitting of cleaned curve using Fourier analysis Least-square fitting to good quality values

Eliminate all short-term variations

Uses NDVI quality information

Iterative fitting of cleaned curve using

Fourier analysis

Least-square fitting to good quality values

Methodology As required by the ARD algorithm, each input and the hidden output is a weights class with its own α α 0 α c INPUTS : Past NDVI (MODIS 3b42) Previous rainfall (TRMM) Temperature (WorldClim) OUTPUT : 16 day predicted NDVI NDVI t Precipitation (t) Temperature (t) … … w 0 w 1 w 2 NDVI (t-1) NDVI (t-2) NDVI (t-n) w p1 w p2 w p3 w o1 w o2 w o3

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.

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.

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 ïƒ 1,150,156,800 vectors ïƒ 23 vectors per year ïƒ 26,453,606,400 NDVI values to manage per year ïƒ 9.5 years of data ïƒ 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

For South America alone, first calculations approximated 10 years of processing for the NN to learn:

A map of 30720 by 37440 pixels

ïƒ 1,150,156,800 vectors

ïƒ 23 vectors per year

ïƒ 26,453,606,400 NDVI values to manage per year

ïƒ 9.5 years of data

ïƒ 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

Sample novelty analysis

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

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

Some statistics 75% of deforestation occurs in December and January 50,000 Ha deforested in Dec/Jan of 2008/2009 compared with 7,500 Ha in 2004/2005 During 16 days of Christmas in 2008 16,000 Ha lost, compared with 500 Ha in 2004 (3%)

75% of deforestation occurs in December and January

50,000 Ha deforested in Dec/Jan of 2008/2009 compared with 7,500 Ha in 2004/2005

During 16 days of Christmas in 2008 16,000 Ha lost, compared with 500 Ha in 2004 (3%)

Parasid Test cases

Introduction Different test cases with different vegetation and climate types All the test are done with the same parameters Training parameters From 2000 to the end of 2003 Detections parameters From 2004 to May 2009 A detection map is created each 16 days within this period The process is near to be fully automated

Different test cases with different vegetation and climate types

All the test are done with the same parameters

Training parameters

From 2000 to the end of 2003

Detections parameters

From 2004 to May 2009

A detection map is created each 16 days within this period

The process is near to be fully automated

Colombia – Río Caquetá Size 480 * 300 [km 2 ] 14400000 [ha] Vegetation type Tropical forest

Size

480 * 300 [km 2 ]

14400000 [ha]

Vegetation type

Tropical forest

Detection : See Caqueta-meta KML See http://www.youtube.com/watch?v=exGmzc70PrQ Pink : Too many clouds to analyse Red : 3 consecutive times detected with more than 95% confidence

See http://www.youtube.com/watch?v=exGmzc70PrQ

Pink : Too many clouds to analyse

Red : 3 consecutive times detected with more than 95% confidence

NDVI 2004.01.01 NDVI 2009.01.01 Anomalies probability 2009.01.01

Colombia – Rio Caquetá

Colombia – Rio Caquetá

Colombia – Rio Caquetá Comments 0.22% deforestation rate per year The model is working well in this area where deforestation seems accelerating

Comments

0.22% deforestation rate per year

The model is working well in this area where deforestation seems accelerating

Colombia – Serranía San Lucas Size 180*960 [km 2 ] 4320000 [ha] Vegetation type Lowland tropical forest Montane forest

Size

180*960 [km 2 ]

4320000 [ha]

Vegetation type

Lowland tropical forest

Montane forest

Colombia – Serranía San Lucas NDVI 2004.01.01 NDVI 2009.01.01 anomalies probability 2009.01.01 Cumulative detection on time

Colombia – Serranía San Lucas

Colombia – Serranía San Lucas Comments 0.63% deforestation rate per year The main area of deforestation in the center is a really strong change which could be flooding. A deeper analysis is needed to explain the different types of changes The model seems to work well even if the place is a bit cloudy

Comments

0.63% deforestation rate per year

The main area of deforestation in the center is a really strong change which could be flooding.

A deeper analysis is needed to explain the different types of changes

The model seems to work well even if the place is a bit cloudy

Colombia – Sierra Nevada Size 120*120 [km 2 ] 1440000 [ha] Vegetation type Dry forest Montane forest

Size

120*120 [km 2 ]

1440000 [ha]

Vegetation type

Dry forest

Montane forest

Colombia – Sierra Nevada NDVI 2004.01.01 NDVI 2009.01.01 Anomalies probability 2009.01.01 Cumulative detection on time

Colombia – Sierra Nevada

Colombia – Sierra Nevada Comments 0.01% deforestation rate per year Difficult to detect changes in this area An analysis of the probabilities is needed to show places with anomalies

Comments

0.01% deforestation rate per year

Difficult to detect changes in this area

An analysis of the probabilities is needed to show places with anomalies

Bolivia – Santa Cruz Size 480*420 [km 2 ] 20160000 [ha] Vegetation type Tropical forest

Size

480*420 [km 2 ]

20160000 [ha]

Vegetation type

Tropical forest

NDVI 2004.01.01 NDVI 2009.01.01 Anomalies probability 2009.01.01

Bolivia – Santa Cruz Cumulative detection on time

Bolivia – Santa Cruz 0.09% deforestation rate

Paraguay - Boquerón Size 240*240 [km 2 ] 5760000 [ha] Vegetation type Savannah

Size

240*240 [km 2 ]

5760000 [ha]

Vegetation type

Savannah

NDVI 2004.01.01 NDVI 2009.01.01 Anomalies probability 2009.01.01

Cumulative detection on time

Paraguay - Boquerón

Paraguay - Boquerón Comments 0.87% deforestation rate Savannah and tropical forest have a totally different environment The model seems to work well even if the changes are more subtle

Comments

0.87% deforestation rate

Savannah and tropical forest have a totally different environment

The model seems to work well even if the changes are more subtle

Chile – Region del Bio Bio Size 240*120 [km 2 ] 2880000 [ha] Vegetation type Tempered forest

Size

240*120 [km 2 ]

2880000 [ha]

Vegetation type

Tempered forest

Chile – Region del Bio Bio NDVI 2004.01.01 NDVI 2009.01.01 Anomalies probability 2009.01.01 Cumulative detection on time

Chile – Region del Bio Bio

Chile – Region del Bio Bio Comments 0.31% deforestation rate The model seems to work with a tempered climate and non-tropical forests

Comments

0.31% deforestation rate

The model seems to work with a tempered climate and non-tropical forests

And now the tough one…

OTCA Amazon Cooperation Treaty Size 4228.75*3498 [km 2 ] 1479216750 [ha] Vegetation type Tropical forest

Size

4228.75*3498 [km 2 ]

1479216750 [ha]

Vegetation type

Tropical forest

OTCA Amazon Cooperation Treaty

OTCA Amazon Cooperation Treaty Comments Average 0.22% deofrestation rate Still a bit noisy in the center Due to clouds undetected during the cleaning process Most of the detections are valid The system seems stable over big areas and a certain amount of consecutive dates (detections over 120 dates)

Comments

Average 0.22% deofrestation rate

Still a bit noisy in the center

Due to clouds undetected during the cleaning process

Most of the detections are valid

The system seems stable over big areas and a certain amount of consecutive dates (detections over 120 dates)

Time processing statistics For an area of the size of OTCA with One Dell server 16 [GB] of RAM 8 processors Intel Xeon X5365 3 [GHz] Cleaning process Cleaning 214 date 12 hours Clustering process 6 Clusters Clustered on the years 2000 to the end of 2003 12 hours Modeling process 3 Models per clusters 2000 pixels as training dataset 5000 pixels as validation dataset 3 hours Detections process for 2004 to 2009 120 detections grids 70 hours Whole process Only 4 days processing from the raw data

For an area of the size of OTCA with

One Dell server

16 [GB] of RAM

8 processors Intel Xeon X5365 3 [GHz]

Cleaning process

Cleaning 214 date

12 hours

Clustering process

6 Clusters

Clustered on the years 2000 to the end of 2003

12 hours

Modeling process

3 Models per clusters

2000 pixels as training dataset

5000 pixels as validation dataset

3 hours

Detections process for 2004 to 2009

120 detections grids

70 hours

Whole process

Only 4 days processing from the raw data

Tasas de deforestación

Model comparison PARASID vs. FORMA PARASID detections First detection in 2004 FORMA probabilities First detection in 2000

PARASID vs DETER It seems Parasid model detects quite small and isolate events which Deter doesn’t detect. 2006 2004

Next Steps Fully functioning web interface January 2010 Preliminary continental validation and calibration (January 2010) Global extent (2011) Additional models to identify type of change (drivers) (2011)

Fully functioning web interface January 2010

Preliminary continental validation and calibration (January 2010)

Global extent (2011)

Additional models to identify type of change (drivers) (2011)

Analysis of three images between the years 2000 and 2009. MATO-GROSSO – BRASIL LAT: - 10.1, LON: - 51.3 10/10/2000 LANDSAT 7 SLC ON 29/06/2009 LANDSAT 7 SLC OFF CLASSIFIED IMAGES IN ERDAS Forest Uncoverage Change 00-09 Unchanged CHANGE DETECTION IN ERDAS

SAMPLING POINTS IN LATIN-AMERICA 1. Covering the whole Latin-America 2. Sampling of different land use type Tropical forest Andes Savanna Desert 3. Selection of areas with high risk of change Near to cities Near to road Near to rivers With crops already existing SELECTION CRITERIA

1. Covering the whole Latin-America

2. Sampling of different land use type

Tropical forest

Andes

Savanna

Desert

3. Selection of areas with high risk of change

Near to cities

Near to road

Near to rivers

With crops already existing

Conclusions Near-real time global monitoring is possible PARASID now functioning for Latin America Providing first approximations of deforestation rates in over a decade for some parts of Latin America

Near-real time global monitoring is possible

PARASID now functioning for Latin America

Providing first approximations of deforestation rates in over a decade for some parts of Latin America

Puntos de mejoramiento Corridas detalladas para zonas Mejoramiento significativo en precisión y cobertura (por nubes) usando datos crudos de MODIS diarios Con calibración y validación en campo Estudio conjunto IDEAM con CIAT/TNC para aproximación nacionales

Corridas detalladas para zonas

Mejoramiento significativo en precisión y cobertura (por nubes) usando datos crudos de MODIS diarios

Con calibración y validación en campo

Estudio conjunto IDEAM con CIAT/TNC para aproximación nacionales

GRACIAS!