Further comment on “Radar backscatter is not a ‘direct measure’ of forest biomass”

As a group of scientists at the universities of Edinburgh and Maryland, and the FAO UN-REDD Programme, we’ve just published a short piece in Nature Climate Change magazine entitled “Radar backscatter is not a ‘direct measure’ of forest biomass”. The focus of this correspondance is “to contest the use of the term ‘direct measurement’ to describe the application of radar backscatter intensity to map forest AGB.”

At first glance, it might appear as a rather negative view of radar remote sensing for forest applications, but we believe its a well-balanced, fair appraisal of where things stand with radar and biomass. We felt it was necessary to counter some of the more enthusiastic proponents of radar for biomass mapping, which we feel has led to a great deal of confusion, especially in the policy domain.

Accurately mapping forest carbon on a global scale is now much more than a scientific question. In the scientific context of planetary global change, mapping forest carbon and forest change has always been high on the agenda. Scientifically, we are still not satisfied with the lack of detail with which we understand how carbon comes and goes between the terrestrial ecosystems and the atmosphere. This puts a limit on the accuracy with which we can predict future climate change.

But forests are no longer just a science issue. The inter-governmental attempts to introduce international policies such as the Reducing Emissions from Deforestation and forest Degradation (REDD+) has put forests at the top of a political agenda. Not only that, but they have mixed it together with a financial model that could see private money as part of the solution (through investment in forest carbon credits, or companies being required to offset their carbon by purchasing carbon credits). And just when you think it couldn’t get more complicated, the International Development agenda also sees sustainable forestry as a fundamental component of livelihood security for many of the poorest communities on Earth.

So, that’s the backdrop to what is starting to look like a land-grab for the world’s forests. And underpinning everything, is the need to measure the state of the forests, and then to report and to verify that measurement (the so-called Measurement, Reporting and Verification procedures, or MRVs).

That’s where radar remote sensing comes in, as its a fantastic tool for measuring forest properties.

Almost three decades of research has demonstrated, without doubt, that the intensity of long wavelength (L-band and larger) radar images is sensitive to forest above ground biomas (AGB), up to some signal-saturation point. Satellite radar is therefore often suggested as the best tool for global-scale mapping for forest biomass, especially since allometric-based field surveys are both difficult and expensive. In the Nature Climate Change piece we clearly promote the idea that radar is typically the best remote sensing tool for mapping forest extent [1], estimating forest structural variability [2][3], and detecting deforestation and degradation [4].

What we take issue with is that in some instances we feel that data are being over- or mis- interpreted, often to match expectations, and this is leading to the case for imaging radar being overstated. My personal fear is that an enthusiastic scientific community, keen to provide a solution to what is an enormous global challenge, are mis-matching a scientific-style solution to a policy-style problem. The risk is that the policy ultimately won’t work if the methods employed aren’t really fit for purpose.

Our specific critique is directed towards the use of the term “direct measurement” [5][6][7] to describe the application of radar backscatter intensity to map forest AGB. It’s easy to slip into this terminology, and I’m sure I’ve done it myself along the way, since you might argue that radar “measures biomass directly” rather than going through some surrogate index. But “direct measurement” implies an unambiguous and well-defined relationship, an assertion that is neither expected from theory nor supported by measurements. I consider it a real worry that reviewers for high impact journals are letting authors away with this kind of language. Even if the paper itself offers a more balanced appraisal in the small print, unfortunately it is the headline items (in the abstract and the figure captions) that can have the biggest influence on non-experts.

Figure 1. Note that both these data sets have essentially two clouds. On the left there is a very low biomass cloud and a high biomass cloud. You can fit any function you like between two clouds of points.
The right hand plot has a cloud of primary forest points and a cloud of secondary forest points. There is little or no correlation in each cloud, but together they provide what looks like a correlation.

Radar backscatter intensity does not provide a direct measurement of forest AGB, even at very long wavelengths ( see this previous blog entry). We still see some fantastic empirical results that continue to demonstrate radar sensitivity to AGB [8] but our understanding of this sensitivity and our ability to generalise across different landscapes is still, in my opinion, a major limitation.

One of the things we couldn’t do in the Nature Climate Change piece (due to page limits) was show some examples of when we believe results are being overinterpreted, so here are some to consider (noting that our intention is not to critique individuals, but rather to show examples of a trends across the community at large). Firstly, we note that some studies present results that cut across different forest types (such as regrowth vs primary forest). In these cases visual examination can clearly differentiates two clusters of points corresponding to different forest types, yet there is clearly no relationship within each cluster. With the application of only simple statistics a high r-squared is still reported, painting an incomplete picture of the data. Figure 1 shows two examples (from published work) where we think this is occuring.

Figure 2. Again, two clouds of points, but note also that the curve that is fitted draws the eye to believe that it is sensitive up to 300+. Indeed, the line does that, but the data doesn’t. In the paper itself they fit the line then draw conclusions from the line, not a visual interpretation, but the casual viewer will see this line and believe the sensitivity is higher than the data suggests (in my opinion). 

Figure 2 shows an extreme example of another analytical error that is common in the literature. A log function is fitted even though a sigmoidal function is predicted by theory, and as a log curve is not asymptotic this visually suggests sensitivity beyond saturation. The line is used to suggest sensitivity up to 252 Mg/ha (for +/-100 Mg/t). Subsequently another fitted log function (not the data), based on X and L band combined, is used to project sensitivity to AGB values higher than 600 Mg/ha, even though there is no field data over 400 Mg/ha. The visually misleading impression caused by drawing a log-line is a common one, and I’ve probably been guilty of that myself. But now I am starting to appreciate that when presented to decison makers who perhaps are not skilled enough in the science to be able to question it, it gives entirely the wrong picture of what is happening in the data.

The real issue that we intended to address in the NCC piece is that many people have been led to believe that radar is a direct measure of forest biomass. In many people’s minds, this has become the default position. I’ve lost count of how many people I have seen write or say something along the lines of, “So, the backscatter-biomass curve is not very good at first glance, so that must be because of this, that and the next thing.” Anything that lies off the well-fitted curve is an “outlier”, and is duly explained away. There is another possible reason why backscatter-biomass curves are not always consistent and well-correlated, and that is because radar is not measuring biomass directly.

Radar is a fantastic tool for forestry, but it is in no-one’s benefit to oversell it.

Figure 3. Anscombe’s quartet. The value of Anscombe’s quartet is that it emphasises the value of looking at data. Each of these four data sets have the same mean, the same fitted line and the same r-squared, yet clearly the data are telling you something very different in each on.
(Image from Wikipedia commons).

[1] Rosenqvist, A., Milne, A., Lucas, R., Imhoff, M. & Dobson, C. A review of remote sensing technology in support of the Kyoto Protocol. Environmental Science and Policy. 6, 441-455 (2003).

[2] Castel, T., Guerra, F., Caraglio, Y. & Houllier, F. Retrieval biomass of a large Venezuelan pine plantation using JERS-1 SAR data. Analysis of forest structure impact on radar signature. . Remote Sens. Environ. 79, 30-41 (2002).

[5] Rosenqvist, A., Milne, A., Lucas, R., Imhoff, M. & Dobson, C. A review of remote sensing technology in support of the Kyoto Protocol. Environmental Science and Policy. 6, 441-455 (2003).

[6] Le Toan, T. et al. The BIOMASS mission: Mapping global forest biomass to better understand the terrestrial carbon cycle. Remote Sens. Environ. 115, 2850-2860 (2011). In the abstract for this paper, the authors state: “During its 5-year lifetime, the mission will be capable of providing […] direct measurements of biomass derived from intensity data…” (my italics).

[7] Englhart, S., Keuck, V. & Siegert, F. Aboveground biomass retrieval in tropical forests: The potential of combined X- and L-band SAR data use. Remote Sens. Environ. 115, 1260-1271 (2011). In the abstract for this paper, the authors state: “The results demonstrate that direct biomass measurements based on the synergistic use of L- and X-band SAR can provide large-scale AGB estimations for tropical forests.” (my italics). What is most surprising is that in the text these authors also state, “no remote sensing instrument can directly measure either biomass or carbon content”. We think there has to be an effort to make sure the language the community uses is consistent.

[8] Carreirs, J.M.B., et al. Understanding the relationship between aboveground biomass and ALOS PALSAR data in the forests of Guinea-Bissau (West Africa). Remote Sens. Environ.121, 426-442 (2012) .



  1. Our figure which is shown in Figure 2 is added with a wrong caption creating false impressions. Our work was misinterpreted or rather not completely understood.
    In contrast to many other studies that can be found in literature, our saturation level is based on a sophisticated calculation and not on visual interpretation.
    Indeed we calculated the saturation on the basis of the fitted curve but is there any possibility to calculate the saturation directly from the data?
    However, the calculation of the saturation level was based on the radiometric accuracy of the radar data and on a chosen accuracy level. We used two accuracy levels (50 and 100 t/ha) and clearly state that “within the accuracy interval of 50 t/ha the estimations are supposed to be accurate whereas estimations within the 100 t/ha accuracy interval are only indicators for the spatial AGB distribution.”
    We found saturation levels of 80 t/ha (TerraSAR-X), 130 t/ha (ALOS PALSAR) and 300 t/ha (TerraSAR-X and ALOS PALSAR) at the 50 t/ha accuracy level. And we emphasized in the discussion that the 100 t/ha accuracy saturation levels are “not accurate enough for a reliable AGB estimation, but can be used for indicating the spatial distribution”.
    Furthermore, there is almost no difference between a log and a sigmoidal fitted function as the slope of the log function is very small in higher biomass ranges.

    1. This is a very important discussion point that the community needs to agree on, namely the fitting of the curve. For instance, we have to decide if the data is saturating or not. If we believe the data is saturating, then we should fit a curve that saturates (sigmoidal or otherwise). If we fit log functions, we are essentially saying the relationship is linear (with real number NRCS). Sometimes the data may suggest that. Other times it clearly doesn’t. Do we know why these cases are different? I’m not convinced we do (although I have blogged my thoughts on this in a previous post).
      My personal concern is that we have reached as stage where we are applying statistics “blind” — meaning, we don’t actually understand what is happening in the data, and we are letting the stats lead our thinking, rather than insight and understanding. Certainly your methodology is rigorous and consistent with many other studies, and we chose it as an example exactly because it was just that.

      On the topic of “spatial distribution”, I’m not sure what you mean? Our concern in the NCC paper is that the radar community (all of us, over the last decade) are responsible for giving an impression of performance beyond the evidence. For example, as you point out, your paper says not to trust the data beyond 300t/ha, yet you show maps of biomass with keys that go to >600t/ha. Decision makers will see the maps, ignore the small print, and be left with an impression that you can go to 600t/ha. I personally believe that, long term, this will be a problem (if its not already).

      By the way: Apologies if the caption implies that you used a visual interpretation. That wasn’t my intention and I will change it again (so if you are reading these comments after this post, note that the caption may have changed).

      1. I’m pleased to note that both Englhart and the BIOMASS mission team have published correspondences in the recent copy of Nature Climate Change. I think this is an important topic that needs discussed. Perhaps they will make pre-publication copies available for everyone to see, and maybe even add links here.

  2. Pingback: Forest Planet
  3. Thanks for such an informative discussions. I am very new in Radar remote sensing and working in a project “Forest parameter characterization using SAR techniques” I read many papers still a confusion in my mind how forest biomass is directly related to backscattering coefficient and what is advantage of microwave data over simple optical data?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s