This research was supported by Natural Environment Research Council through the National Centre for Earth Observation and carried out at the School of Mathematics University of Sheffield

Background

Clearance of the world's primary tropical forests causes a very significant annual loss of biomass to atmospheric carbon dioxide and the destruction of irretrievable biodiversity. Yet there is huge uncertainty in the estimates of global deforestation rates [1]. Reducing this uncertainty is crucial to assessments of global carbon balance for climate modeling and harnessing political will for change.   The economic motivation for deforestation is strong and any change in behaviour is unlikely unless incentives are given to governments, landowners and communities.  Current efforts to do this through the Reduced Emissions from Deforestation and Degradation (REDD) mechanism [2] are dependent on reliable, independent estimates of deforestation rates, undistorted by the vested interests of governments and other parties. Reliable mapping of the changes in tropical forest is therefore crucial to implementing this initiative. Synthetic aperture radar (SAR) images are unaffected by cloud cover and potentially provide an excellent system for measuring forest change in the tropics, as long as they can give sufficiently accurate estimates of deforestation and forest degradation.

Project Aims

The Sheffield project finished in August 2011 but our hope is that these methods will eventually be extended to the whole of Indonesia and provide the Indonesian and global community with a tool to track natural forest cover change as a basis for action on biodiversity conservation and forest carbon management. A paper was published in 2012 [3] detailing much of the work and a further paper describing an enhanced method using texture metrics was well advanced when the project terminated but remains unpublished[4]. Further information, including notes on the methods used and freely available computer code, is given here .

Brief description

In this case study for the Riau and Jambi areas of Sumatra we developed methods to detect deforestation using ALOS-PALSAR data. Two types of data were acquired from JAXA. Fine Beam Dual (FBD) dual-polarised images with a resolution of 12.5 m were used to generate intensity[3] and textural-based[4] change measures. These were compared and combined with detections obtained from 100 m resolution multi-temporal ScanSAR images obtained at 46-day intervals, and assessed using forest change data provided by WWF. Further details are given below and also in this poster.

FBD images

FBD images, covering 70km in range and 59km in azimuth, showing the HH channel in red, HV in green and HH/HV in blue. As forest canopy consists of many scatterers with different orientations it tends to depolarise the scattered signal - hence green often indicates forest cover. Comparing the two figures will show that a number of green areas have changed over the intervening year - often as a result of deforestation.

Detection by intensity ratio

A composite function of the intensity ratios, R1, then combines positive and negative changes. Comparison of probability density distributions (pdf's, left figure) shows that there is significant separation between values of R1 for deforested areas and undisturbed forest. Deforestation can therefore be detected by applying a threshold to select high values of R1. Integrating these distributions gives the probabilities Pd and Pfa for true and false detection of deforestation as the threshold changes. These are plotted (right) as Reciever Operating Characteristic (ROC) curves. The R1 values for HH and HV channels can both detect ~50% of the deforestation for a 20% false alarm rate, but each channel (green) detects significantly different regions, so combining them (by simple summation) as a single measure, R1 (red), improves performance. Textural measures (based on local spatial variation of intensity) behave similarly and can be combined wiith the basic intensities to further enhance detection performance[4]. The ROC curves in the figure below show that detection rates can thus be boosted to ~80% for a 20% false alarm rate.

FBD detection maps

Detections (left figure), using combined intensity and texture measures showing the highest 10% values in red, the top 10-20% values in green and the top 20-30% values in blue; undisturbed forest is shown in white. Regions shown as deforested by the databases (right) are shown red in the right hand figure for the same region.



ScanSAR detection map and combination with FBD detections

Multi-temporal detection using the temporal standard deviation, SD, values for 12 ScanSAR images, were resampled to the same FBD footprint region and can also detect change due to deforestation by applying thresholds to capture the highest values. The figure on the left shows the highest 10% values of SD in red, the top 10-20% values in green and the top 20-30% values in blue with undisturbed forest shown in white. The figure on the right uses the same thresholding scheme for a principle component combination (PCA) of FBD intensity, texture and ScanSAR SD that gave the best detection characteristics of all. Compare with the database derived regions in the figure above.

Comparison by ROC curves

The ROC curves for several individual (open symbols) and combined (closed symbols) measures are compared in this figure. R1 is a measure obtained from the intensity ratios between early and later images, T2, similarly from the texture[4], and SD is the ScanSAR multitemporal standard deviation. For the combined measures sums refers to range-scaled data fusion summation and PCA1 to the first component of a principle component combination of R1, T2 and SD.

References

  1. Achard, F., H. D. Eva, et al. (2002). "Determination of Deforestation Rates of the World's Humid Tropical Forests." Science 297: 999-1002
  2. Parker, C., et al. The Little REDD+ Book. An updated guide to governmental and non-governmental proposals for reducing emissions from Deforestation and degradation. 2009; Available from: Global Canopy
  3. Martin Whittle, Shaun Quegan, Yumiko Uryu, Michael Steüwe, Kokok Yulianto. Detection of tropical deforestation using ALOS-PALSAR: A Sumatran case study. Remote Sensing of Environment 124, (2012), 83-98. Abstract.
  4. Martin Whittle et al. Texture-enhanced detection of tropical deforestation using ALOS-PALSAR. Abstract.

Back to home