Remote Sensing of Forest Degradation in Southeast Asia – Regional Review
Jukka Miettinen, Hans-Jürgen Stibig, Frédéric Achard, Andreas Langner, Silvia Carboni
Institute for Environment and SustainabilityEuropean Commission Joint Research Centre
Presentation outline:
1. Remote sensing in forest degradation monitoring
2.Forest degradation monitoring in Southeast Asia with 5-30 m resolution satellite data
• Continental Southeast Asia• Insular Southeast Asia
3.Lessons learnt
4.Way forward
Remote sensing in forest degradation monitoring
Remote sensing as part of the bigger picture
• Remote sensing can only be a part of forest degradation monitoring and the decisions which data/approach to use is effected e.g. by:
a. National circumstances, particularly geography and types of forest and forest degradation
b. Integration to other aspects of national forest monitoring:i. Satellite data already in useii. Integration to NFIiii. Sampling versus wall-to-wall coverageiv. Availability/type of field data
• Available resources:• Available satellite data• Available budget resources• Hard and software resources• Required training
What do we want to see?What can we see with remote sensing?
• Spatial aspect:
o Types of disturbance (selective logging, fire, shifting cultivation, etc.)o Geography (topography, climate etc.)
• Temporal aspect:
o To verify degradation, the detected disturbance should be evaluated against desired baseline.
o Furthermore, signs of disturbance often disappear rapidly.
It is important to remember that the methods discussed in this presentation primarily look at
indicators of disturbance (potential degradation).Field data/knowledge always needed!
Spatial aspect: Monitoring degraded forests with multi-resolution satellite data in Tanzania
Very High Resolution WV2 data 0.5m
Tree height map derived from WV2
Rapid Eye data 5m
Landsat data 30m
With permission from Dr. Hugh Eva – EC JRC
Temporal aspect:
With permission from Prof. Dr. Florian Siegert – GeoBio-Center & RSS Remote Sensing Solutions, GmbH
Monitoring selective logging with RapidEye data in Myanmar
Selection of satellite data for forest degradation monitoring in national and regional context
Plus: LiDAR (i.e. laser scanning, airborne for degradation monitoring)
Type of data
Example sensors Characteristics
5-30 m spatial resolution
Landsat TM, ETM+, OLI SPOT HRV, HRVIRALOS AVNIR-2, PALSARNational satellitesSentinel-1, -2 (2015->)RapidEye
+ Cheap (or free) data+ Feasibility of time series+ Long historical record+ Large spatial coverage
- Limited spatial resolution
< 5 m spatial resolution
SPOT 6, 7GeoEyeWorldViewTerraSAR-X
+ Improved spatial resolution
- Expensive- High processing capacity demand- Difficult automated interpretation
Forest degradation monitoring in Southeast Asia with 5-30 m resolution satellite data
Continental Southeast Asia – Setting the scene
• Long history of human impact in the form of both selective logging and shifting cultivation, accompanied with fire activity.
Difficult to produce a definition of degraded forest
• Wide range of natural forest types; from fully evergreen closed canopy forests to open canopy dry deciduous savanna with strong seasonal variation in canopy appearance.
Canopy openness not necessarily a sign of disturbance
• Remaining forests largely in mountainous terrains
Topographic issues in remotely sensed data
Forest cover (green) in continental Southeast Asia 2000 (Stibig et al. 2007)
Steepness of the terrain, derived from 90m SRTM (Jarvis et al. 2008)
Concentration of forest on steep landscapes
Continental Southeast Asia – methods used
AuthorsPrimary dataset
Coverage (km2)
Methodology used in degradation analysis
Forest type
Mon et al. 2012b
Landsat ETM+
~2700 Comparison of three semi-automated classification methods to derive canopy cover percentage.
Deciduous
Häme et al. 2013a
ALOS AVNIR and PALSAR
~22 000
Unsupervised fuzzy classification with a sample of very high resolution data (2-4 m) to support medium resolution mapping.
Evergreen to deciduous
Messerli et al. 2009
Existing LC maps
~237 000Land cover mosaic analysis to detect the extent and intensity of shifting cultivation.
Evergreen to deciduous
Potapov et al. 2008
Landsat TM and ETM+
~2 200 000 (entire sub-region)
Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary datasets (e.g. roads).
Evergreen to deciduous
Nov 2013 Nov 2014 NBR difference Google Jan 2014
Disturbance monitoring by vegetation index changes (I)• Bolikhamsai, Lao PDR• NBR difference, Landsat 8 NBR = (NIR-SWIR)/(NIR+SWIR) = [Landsat 8] (b5-b7)/(b5+b7)
NBR = (NIR-SWIR)/(NIR+SWIR)
• Bolikhamsai, Lao PDR• NBR difference Nov
2013 –Nov 2014, Landsat 8
2013 2014
2014 NBR difference 2013-2014
Disturbance monitoring by vegetation index changes (II)
Landsat 8 multispectral (RGB:754)
Landsat 8 pan-sharpened natural colourPan-sharpening for visual interpretation
Insular Southeast Asia – Setting the scene
• Nearly all forest types in insular Southeast Asia have closed canopy in natural undisturbed state.
Canopy openness a good indicator of forest disturbance
• Seasonal variation in forest characteristics is minimal.
No issues with temporal image selection
• Difficult climate conditions for optical remote sensing.
Very limited number of available images, often with remaining atmospheric disturbances
• Remaining forests largely in mountainous terrains
Topographic issues in remotely sensed data
Forest cover (green) in insular Southeast Asia 2000 (Stibig et al. 2007)
Steepness of the terrain, derived from 90m SRTM (Jarvis et al. 2008)
Concentration of forest on steep landscapes
Insular Southeast Asia – Methods used
Authors Primary dataset Coverage (km2) Methodology used in degradation analysis
Langner et al. 2012
Landsat ETM+ ~830 Degradation index taking advantage of SWIR reflectance.
Franke et al. 2012
RapidEye ~4700Spectral mixture analysis followed by decision tree classification.
Miettinen and Liew 2010
SPOT 4 and 5 ~130 000 Visual detection of signs of anthropogenic disturbance.
Wijedasa et al. 2012
Landsat TM and ETM+
~200 000 (flat peatlands)
Maximum likelihood classification with post classification compositing.
Bryan et al. 2013
Landsat TM and ETM+
~200 000 Visual logging road detection and buffering.
Carlson et al. 2013
Landsat TM and ETM+
~320 000 (excl. steep landscapes)
Spectral mixture analysis followed by decision tree classification.
Margono et al. 2012
Landsat ETM+ ~450 000Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Shearman et al. 2009
Landsat and SPOT ~460 000 Visual logging road detection and buffering.
Gaveau et al. 2014
Landsat MSS, TM and ETM+
~740 000 Visual logging road detection and buffering.
Margono et al. 2014
Landsat ETM+ ~1 900 000Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Potapov et al. 2008
Landsat TM and ETM+
~2 700 000 (whole sub-region)
Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Authors Primary dataset Coverage (km2) Methodology used in degradation analysis
Langner et al. 2012
Landsat ETM+ ~830 Degradation index taking advantage of SWIR reflectance.
Franke et al. 2012
RapidEye ~4700Spectral mixture analysis followed by decision tree classification.
Miettinen and Liew 2010
SPOT 4 and 5 ~130 000 Visual detection of signs of anthropogenic disturbance.
Wijedasa et al. 2012
Landsat TM and ETM+
~200 000 (flat peatlands)
Maximum likelihood classification with post classification compositing.
Bryan et al. 2013
Landsat TM and ETM+
~200 000 Visual logging road detection and buffering.
Carlson et al. 2013
Landsat TM and ETM+
~320 000 (excl. steep landscapes)
Spectral mixture analysis followed by decision tree classification.
Margono et al. 2012
Landsat ETM+ ~450 000Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Shearman et al. 2009
Landsat and SPOT ~460 000 Visual logging road detection and buffering.
Gaveau et al. 2014
Landsat MSS, TM and ETM+
~740 000 Visual logging road detection and buffering.
Margono et al. 2014
Landsat ETM+ ~1 900 000Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Potapov et al. 2008
Landsat TM and ETM+
~2 700 000 (whole sub-region)
Visual analysis of signs of anthropogenic disturbance combined with GIS analysis of auxiliary data (roads etc.).
Pixel level detection Segment level aggregation
BC = broken canopy
• Disturbance detection on pixels, followed by aggregation by segments.
Forest disturbance detection with vegetation/soil indices
SELECTIVE LOGGING FOREST FIRE
Shade Fraction ImageLandsat ETM+ Soil Fraction Image Landsat ETM+
Utilization of spectral unmixing in South America (I)
With permission from Dr. Yosio Shimabukuro – European Commission Joint Research Centre
Threshold of soil fraction on Landsat 5 TMSoil Fraction Image
With permission from Dr. Yosio Shimabukuro – European Commission Joint Research Centre
Utilization of spectral unmixing in South America (II)
Lessons learnt
Difficulties, difficulties, difficulties…
• Both continental and insular Southeast Asia have challenging conditions for remote sensing based forest degradation monitoring:
o In the continental Southeast Asia difficulties are mainly related to(1) the variety of natural forest types (e.g. canopy openness)(2) the seasonal variation of reflectance in the satellite data
o In the per-humid Southeast Asia the limited availability and poor quality images is the main source of difficulties.
• Throughout the region, steepness of the landscape complicates forest degradation monitoring by remote sensing.
Why has it been difficult to apply automated methods to large areas?
• Limitations documented in successful small scale studies:
1. The extent of natural forest area should be known, in order to put any subsequent findings in a focused context.
2. Image-by-image sample area/end-member collection often needed.
3. The successful studies have been typically performed on relatively flat and homogeneous forest areas.
4. Empirically defined thresholds of parameters/indicators have been used to define categories for degradation.
5. Automatic filtering or manual clean-up of the results has often been used.
Practically all forest degradation mapping efforts over large areas in the region to-date have been performed using visual interpretation.
With these limitations…
So what CAN we do with 5-30 m spatial resolution data?
How to derive internationally comparable forest disturbance estimates by remote sensing – food for thought…
• Suitable indicator (e.g. SMA, indices etc.) need to be selected.
• A benchmark forest extent is needed, followed by change detection within the known forest area.
• Detected changes can be grouped to coarse strata, e.g.: 1. Deforestation2. Potential disturbance
• Potential work flow: 1. Pixel level change detection 2. Assignment of change strata by pixels or by segments 3. Visual validation
• A coarse stratification scheme (2-3 strata) for derivation of comparable estimates throughout the region. Finer analysis within strata for local / national forest degradation assessments.
Is it possible to apply this type of concept into practice?
• Limitations for time series analysis/change detection have been largely technical (i.e. lack of adequate data and processing capacity).
• Landsat 8 and Sentinel-2 satellites will dramatically improve data availability in Southeast Asia.
o < 5-day observation frequency will allow change detection also in the humid insular part, and improve the quality of available data for the continental part.
• Processing capabilities improving rapidly: Google Earth Engine, CLASlite, IMPACT tool, ImgTools, etc.
• The challenge in the next few years will be to implement regionally robust and consistent approaches to assess and operationally monitor forest disturbance/degradation throughout the region for the upcoming REDD+ era.
• The near future will tell whether the highly anticipated increase in satellite observation frequency and the expanding image processing capabilities (Google Earth, CLASLite, IMPACT tool etc.) will enable monitoring of forest degradation consistently and accurately in the whole region of Southeast Asia.
• At the same time, increasing amount of ≤ 5 m data (RapidEye, SPOT 6 and 7) as well as new approaches of incorporating radar and LiDAR data will create new possibilities (with new processing challenges).
The way forward
Thank you for your attention!
References
The presentation is mainly based on:
•Miettinen, J., Stibig, H.-J. and Achard, F. (2014) Remote sensing of forest degradation in Southeast Asia - aiming for a regional view through 5-30 m satellite data. Global Ecology and Conservation 2: 24-36.
Other recent SEA forest monitoring related publications from JRC:
•Stibig, H.-J., Achard, F., Carboni, S., Raši, R. and Miettinen, J. (2014) Change in tropical forest cover of Southeast Asia from 1990 to 2010. Biogeosciences 11: 247-258.
•Miettinen, J., Stibig, H.-J., Achard, F., Hagolle, O. and Huc, M. First assessment on the potential of Sentinel-2 data for land area monitoring in Southeast Asian conditions. Asian Journal of Geoinformatics 15: (1) 23-30.
•Stibig, H.-J., Carboni, S., Miettinen, J. and Gallego, F.J. Uncertainties in mapping tropical forest cover from satellite imagery of medium spatial resolution– a case study on the TREES results in Southeast Asia. IJRS (in review).
Top Related