Wealth Accounting and the Valuation of Ecosystem Services (WAVES)
Recently, I participated in a training organized by consultants in charge of estimating timber volume in Madagascar using remote sensing techniques.
First step of their work consisted of collecting and processing large landsat images (36 in total covering period 2005 to 2013).
Output of the studies are:
They applied unsupervised classification to form homogeneous group of pixels corresponding to similar geographic entities. Examples: woodland, dense humid forest, degraded humid forest, other vegetation, etc. It helps also reduce the size of images.
At the same time, manual labelling and regrouping of different clusters are performed which reduce the number of clusters at the end. It must also be noticed that at this stage, satellite images have been reduced into a single band due to the classification process. When clusters are clearly identified as corresponding to the different forest classes, they proceed to estimation of area using GIS software.
One issue that has been raised during the training is how to estimate area hidden by cloud as Landsat images are not free of cloud. Consultants proposed 03 cross-checking approach:
The first is to use expert knowledge who knows the terrain . Second is to use Google earth to check what is actually above cloud by searching their geographic coordinates and last, find an image without cloud (this last approach is far from possible most of the time).
After obtaining homogeneous classes by image. They proceed to ensemble all images. It's an important step as Landsat images are segmented into different parts. Mosaicing consists of blending several images to form one large radiometrically balanced image so that the boundaries between the original images are not seen (the process is automatically done by the software).
Field survey comes from IEFN 1995 (Inventaire Ecologique Forestier National ~ national ecologic forest inventory). In fact,IEFN methodology relied on Stratified Random Sampling. Forest area have been stratified into homogeneous group (based on pixels density into small-medium-high classes) using satellite image then crosstabulated with forest kinds to get strata .With the help of GIS tools, they calculated area for all strata. An intensity sampling of 1% of the total area is used and multiplied by each strata size to get to area to be surveyed per strata. Depending on the area per strata, a decision is made on which plot size [XxY] (in meter per square) is suitable for by taking into account species available in each strata.
Data points from this survey are geolocalized into maps (processed by consultants) to estimate volume with the help of linear model(s). Pixel density or Normalized Difference Vegetation Index (NDVI) are used as features.
In fact, It should always exist a linear relationship between volume obtained from field survey (IEFN in this case) and density pixel of the map where data points have been plotted. Consultants exploited this linear relation to build a linear model, then after, predict timber volume on uncovered pixel area. At the end, they covered all Madagascar and obtained a complete estimation of timber volume for the different forest classes. However, sometimes, pixel density may not exhibit linear correlation with timber volume collected. In this case NDVI is used as features to replace density pixel in the linear model.