Please find below, comments on the FAO Voluntary Guidelines on National Forest Monitoring provided to me by Dr. Steen Magnussen.
Best regards, Jeff.
Pretty good document. Below are some specific comments.
1. In section 4.1 the term ‘landscape view’ is used. Meaning what? I have an idea but don’t want to just guess.
2. In the third bullet of 4.1.1 something is missing. A rewrite is in order. In the same section a rather narrow view of the role of RS is put forward (classification).
3. Box 4. There will never be a scientifically based minimum level of acceptable precision. How could there be?
4. In 4.2.1 fourth bullet. Meaning is unclear. The term ‘on-the-ground information’ in the 7th bullet is too vague.
5. If the population frame is determined from a map (say a forest mask) then errors in this frame must be mentioned. Some sample points may actually be outside the frame and should therefore, in theory, be withdrawn from the list of sample locations. However, over time the forest frame may change. Therefore if forests is the population of interest it is best to use ‘all land’ for selection of sample locations but only sample field data on forested locations. Periodic updates of the forest mask will provide important statistics on forest area. Many national inventories apply this approach to avoid otherwise serious issues of a sliding population frame.
6. Box 7 (p. 23) and Box 9. The issue of autocorrelation is only relevant for the efficiency of a sample design. This comes out at a later point. But here it may not be appreciated. Perhaps rephrase the paradigm of optimizing the within plot variability towards minimizing the among-plot variance, which, of course, is done by maximizing the within plot variance.
7. Page 23 mentioned that it is not possible to optimize the sample size for several sample sizes. I think the statements needs to be softened a bit. I can think of solutions if you don’t require the same precision for all variables; and yet others based on optimization given constraints on budget and or time.
8. In the 7th bullet in 4.2.3 it may be worthwhile to mention that the fidelity of geo-referencing in RS is not perfect. In some years, and some locations it is ‘user beware’.
9. In the 8th bullet in 4.2.3 Your dismissal of fixed-count sampling is too harsh. Of course, fixed-count estimators are not design-based. Steen Magnussen has developed new and approximately unbiased fixed-count estimators of stem density and is on the verge of submitting a manuscript illustrating the same desirable property for density estimators of VOL and BA.
10. What is a statistical model? A model is a model. It may be used in connection with a statistical estimation procedure. But the term is foreign to me.
11. On page 28 and forward the term ‘introduces variability’ is used as if variability is exchangeable for errors and even bias. A thorough revision is called for. One that eliminates the ill-defined ‘variability’.
12. On page 30 the sentence ‘when reporting only’ needs a revision. I also found the implicit linking of errors to failures rather mysterious.
13. In 4.2.7 on the issue of QA/QC a statistical approach to quality control should be promoted rather than a rule of thumb (10% of field plots to be remeasured!). In particular a sequential QA/QC statistical analysis is efficient. Remember, a decision has to be made: accept or reject the work. So clear limits of acceptable ‘errors’ or ‘divergence’ are needed.
14. A point worth mentioning? If confidence intervals are given for a large number of inventory attributes they are merely point-wise intervals. One must recognize that the multivariate confidence level is lower. This is important because NFMS deal with many reportable attributes and when deciding on sample size the multivariate nature of the statistics is often ignored.
Director, Forest Information / Canadian Forest Service
Natural Resources Canada / Government of Canada