Comments on ``Failures in detecting volcanic ash from a satellite-based technique''

Fred Prata1, Bill Rose2, Dave Schneider3, and Andrew Tupper4

1CSIRO Atmospheric Research
Aspendale, Victoria, Australia.

2Department of Geological Engineering and Science
Michigan Technological University, Houghton, MI, USA.

3Raytheon STX, U. S. Geological Survey
Alaska Volcano Observatory, Anchorage, Ak, USA.

4Bureau of Meteorology
Darwin, NT, Australia.

10 May 2000

Introduction

The recent paper by Simpson et al. (2000) on failures to detect volcanic ash using the `reverse absorption' technique provides a timely reminder of the danger that volcanic ash presents to aviation and the urgent need for some form of effective remote detection. The paper unfortunately suffers from a fundamental flaw in its methodology and numerous errors of fact and interpretation. These errors are discussed in this note.

The basic premise of the paper is to demonstrate to its audience that a particular satellite-based algorithm (the T4 - T5 method), widely used for detecting hazardous volcanic ash clouds, often fails. The paper also notes that the algorithm is fundamentally incapable of providing prompt detection of the exlposive event itself. The fact that the algorithm fails under certain conditions has been known for a long time (Prata, 1989a,b; Prata and Barton, 1994; Rose et al, 1995) and most of the reasons for failure are already well-known amongst the research and aviation community. The fundamental incapability of the algorithm to detect early ash cloud events, while possibly true, is not shown by the authors. To understand how the authors have reached their conclusions it is worth scrutinising their methodology.

They must show that, against some independent `truth' concerning the existence or non-existence of volcanic ash in a plume (they deal exclusively with plumes and ignore volcanic clouds and highly dispersed ash layers), the T4 - T5 method misclassifies pixels within the plume. The misclassification can occur in two modes: the algorithm can miss pixels that are known to be ash, or it can classify pixels as ash, that are known not to be ash. Either way the authors must demonstrate that they know the true nature of the pixel under question. Their methodology is:
(1) Using the same satellite data, the authors use an arbitrary threshold on pixel brightness temperature or manually determine the location of a plume in the image.
(2) They assume, without justification, that all pixels in the plume are volcanic ash.

The authors have unwittingly presented an alternate algorithm for detecting ash in plumes-one that they assume is 100% perfect. At the most pedantic level it could be argued that neither do they know the truth nor can they assume that the pixels in a plume are volcanic ash. It is also not good practice to compare algorithms using the same data-it is not an independent test. Thus at the start we can see that the basic methodology has some serious problems and certainly should not be used to invalidate a second approach.

We have studied the examples presented in their paper in some detail and illustrate the basic problems with their methodology below.

Analysis

(i) Soufriere Hills (Montserrat).
This eruption occurred in the tropics with quite high amounts or precipitable water (4.8±0.5 cm-quoted in inches in their paper which we have converted to cm). Figure 3 of their paper shows several time-frames of GOES imagery obtained during eruptive activity. The plume is identified perfectly by the authors and surrounded by a rectangular box on the imagery. The actual location of the plume is anything but clear and perhaps the only general agreement about the imagery is that there appears to be cloud-like features near to the island of Montserrat. These could be volcanic in origin. They are not identified as volcanic by the T4 - T5 method and the authors conclude up to 99% false classification rates for the algorithm. Using independent information the authors state that the first image in the series was collected 3 minutes after an eruption which took an ash column up to 13 km ASL. The next image is 3 hours later and we must assume that the cloud over Montserrat is a volcanic plume generated by the eruption 3 hours and 3 minutes previously. High concentrations of silica-rich ash is known to cause jet turbine engines to stall (see Casadevall, 1994). The amount of ash in this plume is not discussed by the authors and it is likely that it was not known. The failure of the T4 - T5 method cannot be assumed because the certainty of ash in the plume cannot be guaranteed. Likewise, the hazard to aircraft therefore cannot be ascertained.

(ii) Mt Spurr/Crater Peak
This eruption occurred in arctic conditions in a dry atmosphere (0.64 cm of precipitable water according to Simpson et al.). Data used are from the AVHRR-2-a superior instrument to GOES for this application because it has better digitisation (10 bits vs 8 bits for GOES) and better infrared calibration. Failure rates for this eruption vary from 3% to 89.3%. A close inspection of the spatial pattern of the T4 - T5 negative differences shows that the algorithm identifies the edges of the plumes in all five cases shown. Panels (g)-(h) of Figure 5 are particularly revealing because they show that the so-called failures occur at the centre of the plume. This is anticipated from the theory of the algorithm (see Prata, 1989b) and are in complete accordance with the underlying physics of the algorithm. Regions classified as negative outside the edges of the plume identified by Simpson et al. as being volcanic ash by their algorithm, may or may not be volcanic ash. The `truth' or an acceptable validation protocol approximating truth has not been demonstrated by the authors.

(iii) Mt Augustine
This is also an arctic eruption in a very dry atmosphere. It seems that the authors are arguing that there was sufficient water available from other sources (e.g. snow, juvenile water in the magma chamber) to provide a volcanic source of water to the atmosphere. The analysis for this case is interesting because this example shows the great value of the T4 - T5 method. The crucial image frame is shown as Figure 7b. The Simpson et al. algorithm fails to detect a plume, while the T4 - T5 method identifies a small plume over the volcano vent. Holasek and Rose (1991) report that an eruption ocurred at 20:22 GMT (see Table 5. of their paper), based on local observations (non-satellite). Thus in all likelihood this plume was volcanic and the T4 - T5 method suggests that there was ash in the plume. Holasek and Rose (1991) also suggest that some of the other negative T4 - T5 differences are due to volcanic ash clouds. These regions cannot be tested against the Simpson et al. method, because according to their methodology only plumes are volcanic. There are many parts of this image which show negative differences-and it is evident that the T4 - T5 method is giving results that could be wrongly interpreted. However, the causes for these negative differences are known. For example, at night over clear land surfaces and snow covered surfaces, it is expected that T4 - T5 will be negative due to low level temperature inversions and land surface emissivity effects (Platt and Prata, 1993). At the edges of clouds, negative differences are often found because of misalignments in the fields-of-view of the two channels used to form the difference. Over very cold surfaces negative differences can occur because the calibration of the infrared channels at these temperature is uncertain1

(iv) Ruapehu
This eruption may be classified as mid-latitude with moderate atmospheric water vapour. The authors quote a 1992 reference that the 1996 Ruapehu ash clouds contained large quantities of surface and ground water. While this is clearly in error, the GMS-5 image frames shown in Figure 9(j)-(r) are remarkable in their excellent consistency in identifying the plume. The infrared channels of the GMS-5 instrument are not well adapted to the T4 - T5 method. This is because the data are 8-bit, poorly calibrated and there is significant overlap of the split-window channels which introduces correlation between the channels (Prata and Cechet, 1999). The digitisation error is evident in the frames of Figure 10. Prata and Grant (2000) have studied this eruption in great detail using AVHRR-2 and ATSR-2 data. They conclude that the ease of identification of the ash cloud was in fact due to the absence of water vapour either in the atmosphere or available from within the crater. Earlier eruptions had emptied Crater Lake, changing the style of eruption from phreatomagmatic to magmatic in style (Bryan and Sherburn, 1999). Prata and Grant (2000) also show that the T4 - T5 method identified very thin ash layers in parts of the North Island of New Zealand where ash falls were reported. This is the kind of independent data that is required for careful validation of any objective ash detection algorithm.

(v) Popocateptl
This is another tropical eruption, but in a relatively dry atmosphere. Panels (j)-(r) of Figure 11 show that the T4 - T5 method successfully identifies a small plume extending southwards from the volcano location. There are large areas of cloud-free land and other regions away from the plume that have negative T4 - T5 differences. These occur mostly at night (cf. panels (n)-(r); local times 17:15 to 01:45). Armed with knowledge that there are likely to be T4 - T5 differences over land at night and that the GOES is an imperfect instrument, these misclassifcations are of no great surprise and can be handled easily by a proficient meteorologist. The great utility of the geostationary satellite instruments is of course its high temporal resolution (up to 15 minutes for the GOES). Imagine then, that these static frames are animated and highlighted using the T4 - T5 algorithm and now show a plume moving southwards. Surely this is excellent, current and useful information for volcanic plume detection.

(vi) Rabaul
This last eruption, also in the tropics, is used to demonstrate a catastrophic failure of the T4 - T5 algorithm. Neither AVHRR-2 nor GMS-5 data showed any significant T4 - T5 negative differences for the Rabaul plume. The reason for this has been elegantly shown by Rose et al. (1995) and was due to the ash particles being coated by ice. The source of the water for the copious amounts of ice produced in the plume was thought to derive from sea-water which entered the magma chamber as the caldera collapsed into the sea. We would argue that the failure to detect negative pixels in this case provides strong support for the basic physics of the algorithm. For ice covered particles, radiative transfer theory shows that T4 - T5 should be positive (Prata, 1989b; Yamanouchi et al., 1987). The arch-shaped distribution in the scatter plot (Fig. 13a) is exactly what is expected from the theory. It is also interesting to note that Prata (1989b) calculated a scenario for a volcanic cloud with a high fraction of ice covered particles mixed in with ash particles. His results (see Fig. 3, solid curve of Prata, 1989b) show a striking resemblence to Fig. 13(a) of Simpson et al.

Discussion

The example eruption cases given by Simpson et al., in our opinion do not demonstrate catastrophic or gross (their words) failures in the algorithm. Rather they show that the algorithm is quite robust when used and interpreted in the correct manner. Misclassifications do occur and the reasons for these are known. The radiative transfer theory outlined by Simpson et al. demonstrates a poor understanding of the physical basis for the algorithm. A complete explanation of the physical mechanism for negative differences observed for volcanic ash clouds has been given by Prata (1989b), Prata and Barton (1994) and Wen and Rose (1994). The physics of the problem involves scattering and neglect of this process (as was done by Simpson et al.) yields erroneous results.

We summarise here the conditions under which the algorithm gives negative differences in the absence of volcanic ash plumes.
(1) Over clear land surfaces at night. In the presence of strong surface inversions in temperature and moisture it has been shown by Platt and Prata (1993) that T4 - T5 can be negative.
(2) Clear desert surfaces. Barton and Takashima (1986) demonstrated that negative differences may occur over soils with a high quartz content (e.g. deserts). This is thought to be due to the restrahlen effects mentioned in the Simpson et al. paper.
(3) Over very cold surfaces (temperatures less than 220 K). Two reasons have been noticed for causing negative differences in these conditions. Potts and Ebert (1996) suggest that negative differences occur at the tops of very cold clouds because of overshooting which causes a temperature inversion at the cloud top. However, negative differences also occur over ice covered surfaces (see Yamanouchi et al., 1987) and it is possible that the cause may be due to errors in the nonlinearity corrections used in the calibration procedure (see Steyn-Ross et al., 1992 for a discussion of the AVHRR-2 nonlinearity correction). To assess the importance of these instrumental calibration errors on the T4 - T5, we have looked at AVHRR-2 data on three satellites (viz. NOAA-7, NOAA-9 and NOAA-11) and compared the differences for very cold cloud tops (cloud-top temperatures less than 200 K). Comparison of these data shows that the negative difference as a function of scene temperature, at low temperatures, is siilar for each instrument but different between instruments. This suggests that the effects are instrumental in nature rather than due to cloud and/or atmospheric properties. Thus, we suggest that these negative differences are due to poor instrument calibration (at low temperatures) rather than failure of the technique.
(4) At the edges of clouds. This effect has been noticed in AVHRR data for a long time and is due to misalignment of the centres of the fields-of-view (FOV) of the infrared channels. If the radiance field changes sharply within the instrument FOV then the nonlinearity of the Planck function is sufficient to introduce spurious effects when a difference is taken. The differences can also be very high and positive.

The effects of viewing geometry are also important. Prata and Barton (1994) have shown that the size of the negative difference observed for ash clouds is diminished at high zenith viewing angles (long atmospheric paths). This is due to increased water vapour absorption which introduces positive temperature differences. Simpson et al. did not discuss this effect in their paper. We emphasise that these effects do not constitute failures of the algorithm-they are what is expected from the physics of the problem and users should be aware of them.

Aviation meteorologists are experts in interpreting weather patterns and have used satellite imagery for several decades. Their experience has taught them that the atmosphere is a very dynamic place and that image animation is a very powerful interperative tool. The Australian Bureau of Meteorology rely heavily on high-temporal (30 minutes to 1 hour) information and their Volcanic Ash Advisory Centre (VAAC) in Darwin uses GMS imagery routinely. These data are usually the first to be consulted in an armoury of data that are used to advise of volcanic ash hazards. The T4 - T5 algorithm has been used at the Darwin VAAC since 1994 and the operational meteorologists have gained experience from using it, including occasions when the signal should not be interpreted as a volcanic ash plume. The Darwin VAAC scrutinise the shape of the distribution of pixels in the T4 - T5 vs T4 scatter plot whenever an eruption plume is suspected. This is because, as explained by Prata (1989b) and Wen and Rose (1994), volcanic ash causes a distinct `U' shaped scatter plot, whereas other phenomena cause an `arch' shape in the scatter plot. Even though the differences may be positive, the `U' shape is a fingerprint for ash particles in the atmosphere. Conversely, negative differences ocurring in an arch or other distribution shape are not due to volcanic ash particles. The fact that the Australian Bureau of Meteorology continue to use the algorithm suggests that it is providing useful information. The physical basis for the algorithm (not discussed nor challenged by Simpson et al.) also makes it a powerful method and suggests when the algorithm works best and fails. Barton and Prata (1994) have proposed an airborne instrument based on the T4 - T5 principle. The immediacy of an airborne detection system and the careful control of false alarms makes such a system attractive for volcanic ash cloud avoidance.

To settle the question of the false alarm rate of the T4-T5 algorithm, independent information on volcanic ash concentrations in plumes is required. Schneider et al. (1999) have compared TOMS retrievals of SO2 with ash retrievals from the AVHRR-2 T4 - T5 algorithm for the largest El Chichon eruptions in 1982. Their results indicate a high degree of coincidence between the locations of the volcanic clouds determined by both sensors (see Plate 1. of their paper). They also show that because these algorithms are sensitive to different volcanic constituents (the TOMS to SO2 and the AVHRR-2 to ash) gravitational separation can be inferred when the algorithms indicate different dispersal patterns. Schneider et al. (1999), Bluth et al. (1995), and Krotkov et al. (1999) have shown that TOMS-based SO2 retrievals matched the 2-dimensional dispersal pattern determined by the AVHRR-2 T4 - T5 algorithm for the 1992 Spurr eruption clouds. Finally, Shocker et al. (2000) have used three different sensors and ground-based observations to study the atmospheric residence of Lascar's April 1993 eruption clouds.

Simpson et al. have assumed that they can provide an ash plume detection algorithm which purports to be the `truth' and against which other algorithms can be tested. Readers of their paper should note this and consider their conclusions appropriately.

References

Barton, I. J. and Prata, A. J. 1994, Detection and discrimination of volcanic ash clouds by infrared radiometry - II: experimental. In: Casadevall, T. J., editor. Volcanic Ash and aviation safety: proceedings of the first International Symposium on Volcanic Ash and Aviation Safety; Seattle, Wash. Washington, D.C.: U.S. G.P.O., U.S. Geological Survey bulletin., 2047, 313-317.

Barton, I. J., and T. Takashima, An AVHRR investigation of surface emissivity near Lake Eyre, Australia, 1986, Remote Sens. Environ., 20, 153-163.

Bluth, G. J. S., Scott, C. J., Sprod, I. E., Schnetzler, C. C., Krueger, A. J., and Walter, L. S., 1995, Exlposive emissions of sulfur dioxide from the 1992 Crater Peak eruptions, Mount Spurr Volcanoc, Alaska, U. S. Geol. Surv. Bull., 2139, (Spurr Eruption, Edited by T. Keith), pp. 37-46.

Bryan, C. J., and Sherburn, S., 1999, Seismicity associated with the 1995-1996 eruptions of Ruapehu volcano, New Zealand: narrative and insights into physical processes, J. Volcanol. Geotherm. Res., 90, 1Ü18.

Casadevall, T. J. (Ed.), 1994, Volacnic ash and aviation safety: Proceedings of the First International Symposium on Volcanic Ash and Aviation Safety, U. S. Geol. Surv. Bull., 2047, 450pp.

Holasek, R. E., and Rose, W. I., 1991, Anatomy of 1986 Augustine volcano eruptions as revealed by digital AVHRR satellite imagery, Bull. Volcanol., 53, 420-435.

Krotkov, N. A., Torres, O. , Seftor, C. , Krueger, A.J. , Kostinski, A., Rose, W. I. , Bluth, G. J. S., Schneider, D. J. and Shaefer, S. J., 1999, Comparison of TOMS and AVHRR volcanic ash retrievals from the August 1992 eruption of Mount Spurr, Geophys Res Lett., 26: 455-458.

Platt, C. M. R. and Prata, A. J., 1993, Nocturnal effects in the retrieval of land surface temperatures from satellite measurements, Remote Sensi. Environ., 45:127-136.

Potts, R. J., and Ebert, E. E., 1996, On the detection of volcanic ash in NOAA AVHRR infrared satellite imagery, In 8th Australasian Remote Sensing Conference, Canberra, Remote Sensing and Photogrammetry Association Australia Ltd., Floreat, Western Australia, March, pp. 25-29.

Prata, A. J., 1989a, Observations of volcanic ash clouds using AVHRR-2 radiances. Int. J. Remote Sensing, 10(4-5), 751-761.

Prata., A. J., 1989b, Radiative transfer calculations for volcanic ash clouds, Geophys. Res. Lett., 16(11), 1293-1296.

Prata, A. J. and Barton, I. J. 1994, Detection and discrimination of volcanic ash clouds by infrared radiometry - I: theory. In: Casadevall, T. J., editor. Volcanic Ash and aviation safety: proceedings of the first International Symposium on Volcanic Ash and Aviation Safety; Seattle, Wash. Washington, D.C.: U.S. G.P.O., U.S. Geological Survey bulletin, 2047, 305-311.

Prata, A. J., and Cechet, R. P., 1999, An assessment of the accuracy of land surface temperature determination form the GMS-5 VISSR, Remote Sens. Environ., 67: 1-14.

Prata, A. J., and Grant, I. F., 2000, Retrieval of microphysical and morphological properties of volcanic ash plumes from satellite data: Application to Mt. Ruapehu, New Zealand, Submitted to Quart. J. Roy. Meteorol. Soc.

Rose, W. I., Delene, D. J., Schneider, D. J., Bluth, G. J. S., Kruger, A. J., Sprod, I., McKee, C., Davies, H. L., and Ernst, G. J., 1995, Ice in the 1994 Rabaul eruption: Implications for volcanic hazard and atmospheric effects, Nature, 375, 477-479.

Schneider, D. J., Rose, W. I., Coke, L. R., and Bluth, G. J. S., 1999, Early evolution of a stratospheric volcanic eruption cloud as observed with TOMS and AVHRR, J. Geophys. Res., 104(D4), 4037-4050.

Shocker, H. L., Rose, W. I., Bluth, G. J. S., Prata, A. J., and Viramonte, J. G., 2000, Lascar volcanic clouds of 1993: Merging of satellite-based remote sesning from TOMS, AVHRR and ATSR during three days of atmospheric residence, Submitted to Int. J. Remote Sens.

Simpson, J. J., Hufford, G., Pieri, D., and Berg, J., 2000, Failures in detecting volcanic ash from a satellite-based technique, Remote Sens. Environ., 72: 191-217.

Steyn-Ross, D. A., Steyn-Ross, M. L., and Clift, S., 1992, Radiance calibrations for Advanced Very High Resolution Radiometer Infrared Channels, J. Geophys. Res., 97(C4), 5,551-5,5568.

Wen, S., and Rose, W. I., 1994, Retrieval of sizes and total masses of particles in volcanic clouds using AVHRR bands 4 and 5, J. Geophys. Res., 99(D3), 5421-5431.

Yamanouchi, T., Suzuki, K., and Kawaguchi, S., 1987, Detection of clouds in Antarctica from infrared multispectral data of AVHRR, J. Meteor. Soc. Japan, 65(6), 949-961.


Footnotes:

1 Readers should be aware that at scene temperatures of -60 ¡C, corrections of up to 3 ¡C are added to the 11 and 12 mm channels of the AVHRR-2.


File translated from TEX by TTH, version 2.25.
On 11 May 2000, 15:49.