Near Real Time Monitoring of Habitat Change Using Neural Network and a MODIS data Louis Reymondin – Alejandro Coca – Andy Jarvis Jerry Touval – Andres Perez-Uribe – Mark Mulligan
May 20, 2015
Near Real Time Monitoring of Habitat Change Using
Neural Network and a MODIS data
Louis Reymondin – Alejandro Coca – Andy Jarvis Jerry Touval – Andres Perez-Uribe – Mark Mulligan
Terra-i is a system of habitat changes monitoring that uses different
mathematical models that combine vegetation data (MODIS NDVI) and
precipitation data (TRMM) to detect deviations from the natural cycle of
the vegetation over time and thus antrophogenic impacts on natural
ecosystems.
What is Terra-
It has maps of habitat loss every 16 days at the continental level with 250 meters of spatial resolution.
To use high-frequency imaging and moderatespatial resolution for ...
Monitoring the conversion of natural habitats in near real time. (Results 2
months after the date of capture)
Have a continental coverage of all types of habitat.
Be a support for government agencies in making decisions.
Quantifying habitat conversion rates and make analysis of trends from
2004 to date.
Monitor the impact on protected areas in Latin America.
Terra- goals
Terra-i is a model to predict the evolution of vegetation greenness intensity, based on measures of vegetation behavior in time and current weather measurements to detect significant habitat
changes..
Terra- approach
The intensity of vegetation greenness is a natural cycle that depends on climatic factors (precipitation, temperature), site variables (type of vegetation, soil characteristics) and disturbances (natural or anthropogenic).
Inputs data1. Vegetation Index (MOD13Q1 MODIS Product , 16 days, 250m)
Normalized difference vegetation index (NDVI) represents the amount and vigor of vegetation. In each area the values are closely related to vegetation
type and climatic conditions as well as the predominant land use pattern.
Tiles MODIS level analysis
This gives us greater automation of the process, synchronizing the stages download, pre-processing of MODIS data, Terra-i processing load and soon final
results in the map server and FTP.
Processed Terra-i data Incoming Test Tiles / Terra-i
Input data
2. Precipitation Data of the Tropical Rainfall Measuring Mission(3hours, 28km)
TRMM is led by NASA and the Japan Aerospace Exploration Agency (JAXA). It monitors and studies tropical and subtropical rainfall,
between 35 º N and 35 º S. It was released on November 27th, 1997 from Japan.
The methodology can be split into two main steps:
The training step (using data from 2000 to 2004)• Models are trained in order to find the relationship
between recent precipitation and the changes in the color of the vegetation (for different vegetation types)
The detection step (using data from 2004 to present)• The trained models output are compared with the
satellite measurements in order to detect anomalies in the vegetation state.
Research methodology overview
1.
2.
Model trainingNDVI and QA MODIS data MOD13Q1, Precipitation
(TRMM 3b42)(2000-2004)
Time series gap-filling and smoothing
ClusteringK-Means
Random pixels sampling for each cluster
Neural network training
To reduce the noise present in the data (clouds, atmospheric variations, shadows…)
To reduce processing duration, the NDVI time series with the same trends during the years are grouped together
Cleaned NDVI dataOriginal NDVI data
Time series gap-filling and smoothing
NDVI and QA MODIS data MOD13Q1, Precipitation (TRMM)
(2004-2011)
Difference between the NDVI sensor measurement and the NDVI predicted
by the neural network
Calibration using habitat changes maps generated with Landsat satellite images (30m)
Clasification of change
Vegetation changes maps
Rules
maps of change
probabilities
NDVI increase
NDVI decrease (anthropogenic)Results
NDVI Prediction from 2004 to 2011
Floods
Anomaly detection
Drought
OUTPUT: 16 day predicted NDVI
PredictionMultilayer perceptronBayesian Neural Network (BNN)
Model trainning and noise approximationScaled Conjugate Gradient (SCG)Gaussian noise
Input automatic selectionAutomatic relevance determination (ARD)
The goal of the model is to predict what is the NDVI value at the date t taking as input the NDVI values at t-1, t-2 … t-n and the previous rainfall.
INPUTS: Past NDVI (MODIS 13Q1)Previous rainfall (TRMM 3b42)
change
Methodology – Change detection
Calibration with Landsat Images
As Terra-i generates maps of conversion probabilities, we use Landsat images in order to calibrate the results and select the most appropriate probability threshold for each cluster to generate binary changed/unchanged maps.
2004
2009
Terra-i results comparation with local models
Terra-i results were compared with deforestation data produced by the National Institute for Space Research Instituto Nacional de Pesquisas Espaciais (INPE) from 2004 to 2009 through monitoring systems as PRODES and DETER.
PRODES The Project of estimation of deforestation in the Brazilian Amazon (PRODES) generated estimations from 2003 using a digital classification system with Landsat images (30m).
DETER DETER is a near real time deforestation detection system. It publishes fortnightly deforestation alerts for the Brazilian Amazon using MODIS images (500m).
The comparison shows a high correlation between Terra-i and PRODES systems.
Comparison with PRODESComparison with PRODES
% of PRODES detection within a MODIS pixels
% o
f mat
chin
g de
tecti
ons
The Software
Results
2004 – 2012
Habitat Loss in Colombia 2004-2011
*
Annual Rate : 118,026 Ha/yearTotal Loss: 944,206 Ha
Habitat Loss in The Biological corridor in Meta
Annual Rate: 1,789,138 Ha/yearTotal Loss: 13,418,538 Ha
Habitat Loss in Brasil 2004-June 2011
*
Road impact assessment
The Trans-Chaco Highway (2002-2006), Paraguay
Conclusions
• Very high levels of deforestation pre- and post- road construction
• But > 300% increase in deforestation rates since road finished, with a footprint that
likely goes beyond 50km buffer
Road: Trans-Chaco HighwayProject period: 2002-2006Average pre-road deforestation rate: 23,000Average post-road deforestation rate: 97,000 (+319%)Year of peak deforestation: 2010Footprint (modal deforestation distance): 30-40km
Road impact assessment
The Trans-Chaco Highway, Paraguay
Improve more and more our system by developing methodologies for analyzing the information generated.
Current Projects
Terra -
Deforestation patterns
Potential deforestation at T=0 Potential deforestation at T=150
Predicted deforestation Actual deforestation (Terra-i)
Base map
PROOF OF CONCEPT
Future Deforestation Scenarios
BR-364 Road, Brasil
• Terra-i can also be used within the WaterWorld and Co$ting Nature Policy Support Systems to
understand the impact of recent land cover change on hydrology and the production and
delivery of ecosystem services.
• Data: http://geodata.policysupport.org/
Integration Terra-i with others Policy Support Systems
Water flows Erosion
www.terra-i.org
&“The best way improve a system is to get people to use it”
Dr. Mulligan (Kings College of London)
A mapping and monitoring system for rapid assessment of land cover conversion at a medium scale (250m).
A tool for monitoring conversion of habitat at continental, national and regional level in close to real time.
A tool for understanding the effectiveness of protected areas and other conservation measures in stabilizing or reducing land cover conversion.
A spatial support system for decision making in public policy and private development initiatives.Through its linkage with WaterWorld and Co$ting Nature, a system for understanding the likely impacts of near real-time land cover change on a wide range of ecosystem services.
Conclusions
Terra-i is:
Terra-i isn’t:X Detailed monitoring tool in local level. For this it requires second-level monitoring (with high resolution images) and third level (field data).
X A system to monitor degradation.
Contact us:[email protected]@cgiar.orgwww.terra-i.org