Monitoring the Agung (Indonesia) Ash Plume of November 2017 by Means of Infrared Himawari 8 Data

The Agung volcano (Bali; Indonesia) erupted in later November 2017 after several years of quiescence. Because of ash emissions, hundreds of flights were cancelled, causing an important air traffic disruption in Indonesia. We investigate those ash emissions from space by applying the RSTASH algorithm for the first time to Himawari-8 data and using an ad hoc implementation scheme to reduce the time of the elaboration processes. Himawari-8 is a new generation Japanese geostationary meteorological satellite, whose AHI (Advanced Himawari Imager) sensor offers improved features, in terms of spectral, spatial and temporal resolution, in comparison with the previous imagers of the MTSAT (Multi-Functional Transport Satellite) series. Those features should guarantee further improvements in monitoring rapidly evolving weather/environmental phenomena. Results of this work show that RSTASH was capable of successfully detecting and tracking the Agung ash plume, despite some limitations (e.g., underestimation of ash coverage under certain conditions; generation of residual artefacts). Moreover, estimates of ash cloud-top height indicate that the monitored plume extended up to an altitude of about 9.3 km above sea level during the period 25 November at 21:10 UTC–26 November at 00:50 UTC. The study demonstrates that RSTASH may give a useful contribution for the operational monitoring of ash clouds over East Asia and the Western Pacific region, well exploiting the 10 min temporal resolution and the spectral features of the Himawari-8 data.


Introduction
The Gunung Agung is a stratovolcano located on the island of Bali (Indonesia; see Figure 1) which erupted in late November 2017, after more than 50 years of quiescence (the eruptive event of 1963-1964 was the largest of the twentieth century in Indonesia causing significant damages and many fatalities [1][2][3]). The eruption forced the evacuation of 40,000 people from villages located around the volcano and caused the cancellation of some international flights, leading to a temporary air traffic disruption [4].
In particular, after a number of earthquakes recorded since August 10 2017 the seismicity increased from the end of the same month [5], prompting the PVMBG (Pusat Vulkanologi dan Mitigasi Bencana Geologi), which is responsible for volcanic, seismic, and landslide risk mitigation in Indonesia, to raise the alert level to 2 (on a scale of 1-4). On 13 September, a sulphurous plume rose from the bottom of the crater and five days later an exclusion zone of 6 km was established; the alert level was increased to level 3 [5]. On 22 September, the PVMBG increased the alert level to level 4, expanding the exclusion In this work, we investigate the ash events of 25-28 November 2017 from space by implementing the well-established RSTASH algorithm [14][15][16], which was previously tested over the Asiatic region using infrared MTSAT-1R/2 (Multi-Functional Transport Satellite-1R/2) observations [17], for the first time on Himawari-8 data. In addition, we analyze the short-term temporal fluctuations of ash cloudtop height, evaluating the impact of RSTASH detections on estimates of this parameter.
Although a number of studies up to now have been performed exploiting Himawari-8 observations (e.g., [21][22][23][24]), only a few of them focused on volcanic ash (e.g., [25,26]). Some authors In this work, we investigate the ash events of 25-28 November 2017 from space by implementing the well-established RST ASH algorithm [14][15][16], which was previously tested over the Asiatic region using infrared MTSAT-1R/2 (Multi-Functional Transport Satellite-1R/2) observations [17], for the first time on Himawari-8 data. In addition, we analyze the short-term temporal fluctuations of ash cloud-top height, evaluating the impact of RST ASH detections on estimates of this parameter.

Data
Himawari-8 was launched on 7 October, 2014 and positioned at 140.7 • E [18 -20]. The AHI (Advanced Himawari Imager) sensor, onboard the new Japanese geostationary satellite, has 16 spectral channels ranging from 0.47 µm to 13.3 µm, with a spatial resolution from 0.5-1 km (visible) to 2 km (infrared bands). The AHI performs full disk measurements every 10 min; however, some target areas (e.g., Japan) can be observed with a higher temporal sampling of 2.5 min [20]. The improved features of the AHI instrument, in terms of spectral, spatial and temporal resolution, should guarantee a more efficient monitoring of rapidly changing weather/environmental phenomena in comparison with imagers of the previous MTSAT series [18]. In particular, the AHI infrared bands 7 (3.74-3.96 µm), 11 (8.44-8.76  Although a number of studies up to now have been performed exploiting Himawari-8 observations (e.g., [21][22][23][24]), only a few of them focused on volcanic ash (e.g., [25,26]). Some authors have used, for instance, a qualitative ash RGB product designed by the JMA (Japan Meteorological Agency) [27] to discriminate ash and SO2 plumes from meteorological clouds, emphasizing the advantages of using data from Himawari-8 for monitoring those features in a timely manner [25].

Methods
Algorithms developed to detect ash clouds from space, using sensors like MODIS (Moderate Resolution Imaging Spectroradiometer), generally exploit the reverse absorption effect of silicate particles in the TIR (Thermal Infrared) band at around 11 µm and 12 µm, in comparison with water droplets and ice [28]. The split window technique identifies airborne ash by means of a fixed threshold test applied to difference of Brightness Temperatures (BT) measured at the aforementioned wavelengths, i.e., BT 11 − BT 12 [29]. Advanced detection methods, minimizing the impact of atmospheric water vapor on the above-mentioned brightness temperature difference (BTD) or analyzing signals measured in other spectral bands such as MIR (Medium Infrared) and/or VIS (Visible) ones, have shown a higher efficiency in identifying ash clouds (e.g., [30][31][32][33][34][35][36][37][38]). RST ASH [16] is an ash detection method, based on the Robust Satellite Technique (RST) multi-temporal approach [39], running on both polar [14,15,40] and geostationary [17] satellite data. This algorithm, which does not use fixed and empirical thresholds, has been tested with success in different geographic areas (e.g., [16,41]).

RST ASH Algorithm
The RST ASH algorithm identifies airborne ash by means of two local variation indices defined as: In Equation (1), ∆TIR(x, y, t) = BT 10.4 (x, y, t) − BT 11.2 (x, y, t) when Himawari-8 data are used; BT 10.4 (x, y, t) and BT 11.2 (x, y, t) are the brightness temperatures respectively measured in bands 13 and 14 of AHI sensor (see Section 2), while u ∆TIR (x, y) and σ ∆TIR (x, y) are the temporal mean and standard deviation computed, for the same location (x, y), processing plurennial time series of cloud-free homogeneous (i.e., same calendar month, same overpass time, same spectral channel/s) satellite records. Although bands 13 and 15 of AHI have their central wavelengths close to infrared channel 1 (10.4-1.3 µm) and channel 2 (11.5-12.5 µm) of the MTSAT-2 imager [20], several tests performed on different BTD signals have suggested us to analyze the above-mentioned spectral difference, instead of BT 10.4 (x, y, t) − BT 12.4 (x, y, t), to detect ash.
In Equation (2), BT MIR (x, y, t) − BT TIR (x, y, t) is the difference of the MIR (Medium Infrared) and TIR signals respectively measured at 3.9 µm and 10.4 µm wavelengths (i.e., AHI bands 7 and 13), whereas u MIR−TIR (x, y) and σ MIR−TIR (x, y) have the same meaning as above.
We expect negative values of the ⊗ ∆TIR (x, y, t) index in the presence of ash/dust clouds with particles having a diameter lower than 10 µm [16,42]. The ⊗ MIR−TIR (x, y, t) index, which is generally positive in both nighttime and daylight conditions, is used in combination with the previous one to reduce false positives [16].
RST ASH is capable of guaranteeing performance comparable to water vapor-corrected BTD procedures [43], without requiring ancillary data, despite some limitations widely described in previous works (e.g., [44,45]). Since the algorithm is tunable, we can use different cutting levels of its local Remote Sens. 2018, 10, 919 4 of 16 variation indices for mapping ash-affected areas. We have used the following algorithm configuration to identify ash pixels regardless of illumination conditions: High confidence level : ⊗ ∆TIR (x, y, t)< −3 AND ⊗ MIR−TIR (x, y, t) >0 (3) Mid-confidence level : ⊗ ∆TIR (x, y, t)< −2 AND ⊗ MIR−TIR (x, y, t) >0 Low confidence level : ⊗ ∆TIR (x, y, t)< −1 AND ⊗ MIR−TIR (x, y, t) >0 Moreover, a spatial noise-reduction filter has been implemented within the RST ASH process (in detection phase) to remove sparsely distributed pixels which are generally not ascribable to ash, based on that previously performed using MSG-SEVIRI (Meteosat Second Generation-Spinning Enhanced Visible and InfraRed Imager) data [41].
RST ASH Implementation on Himawari-8 Data To run the RST ASH algorithm, a setup activity devoted to generate the library of spectral reference fields (i.e., temporal mean and standard deviation images), has to be performed. Here only three years of Himawari-8 data, provided by the Center for Environmental Remote Sensing (CEReS) of the Chiba University (Japan), were available for this purpose.
In more detail, after selecting satellite imagery according to the aforementioned homogeneity criteria, an automated procedure was used to: (i) calibrate data; (ii) extract a Region of Interest (ROI) centered over Indonesia from the full Himawari-8 disk; and (iii) remove cloudy pixels from the scenes. Concerning the last step of this procedure, we used the OCA (One Channel Cloud Detection Approach) method [46] to filter out cloudy pixels from the scenes, before generating the spectral reference fields for some nighttime (i.e., 15:00 UTC) and daytime slots (i.e., 22:00 UTC; 23:00 UTC; 00:00 UTC). OCA is a RST-based cloud detection scheme that identifies cloudy radiances based on signal divergence from the unperturbed "clear-sky" conditions; in daytime, signals measured both in the VIS (at 0.6 µm) and TIR (at 10.4 µm) bands are analyzed for this purpose. This cloud screening procedure generally performs well regardless of illumination conditions; i.e., even in nighttime when only the TIR band is used. However, in areas characterized by an almost persistent cloud coverage it may be less effective in filtering meteorological clouds [41,47]. In Figure 2, we show an example of temporal mean and standard deviation images which were generated for the month of interest and for the time slot of 00:00 UTC (i.e., 09:00 JST; Japan Standard Time).
It is worth mentioning that about 100 Gbyte of data were analyzed to perform the setup activity detailed above. Thus, an extensive data archive is required for a full processing of Himawari-8 data (144 time slots in total). In addition, the multi-temporal analysis of satellite records could be time demanding, because of elaboration processes, if proper computational resources are not available. To reduce the amount of Himawari-8 data to be processed, we have used the same spectral reference fields to run RST ASH on contiguous time slots. Table 1 shows, for instance, that the temporal mean and standard deviation images shown in Figure 2 have been used to run RST ASH also at the time slots of 23:40-23:50 UTC and of 00:10-00:50 UTC; i.e., considering a maximum temporal gap (∆t) of 50 min (see third column). We show and discuss results arising from the use of this implementation scheme, which is for the first time applied to geostationary satellite data, in the following sections. in the VIS (at 0.6 µ m) and TIR (at 10.4 µ m) bands are analyzed for this purpose. This cloud screening procedure generally performs well regardless of illumination conditions; i.e., even in nighttime when only the TIR band is used. However, in areas characterized by an almost persistent cloud coverage it may be less effective in filtering meteorological clouds [41,47]. In Figure 2, we show an example of temporal mean and standard deviation images which were generated for the month of interest and for the time slot of 00:00 UTC (i.e., 09:00 JST; Japan Standard Time). It is worth mentioning that about 100 Gbyte of data were analyzed to perform the setup activity detailed above. Thus, an extensive data archive is required for a full processing of Himawari-8 data (144 time slots in total). In addition, the multi-temporal analysis of satellite records could be time demanding, because of elaboration processes, if proper computational resources are not available. To reduce the amount of Himawari-8 data to be processed, we have used the same spectral reference fields to run RSTASH on contiguous time slots. Table 1 shows, for instance, that the temporal mean

Ash Cloud Detection and Tracking
In Figure 3, we show two RST ASH maps generated from Himawari-8 data of 26 and 28 November 2017, by using the ash detection scheme previously described in Section 3.1. Maps display ash pixels in three different colors based on different confidence level of detection (i.e., yellow: low; orange: mid; red: high; see Section 3.1), providing qualitative information about the areas affected by ash.
In detail, Figure 3a shows the RST ASH product of 26 November at 00:00 UTC (09:00 JST), revealing the presence of an ash plume within the area marked in black and magnified on the top right side of the figure. The plume, moving from Mt. Agung (indicated in purple) towards SE direction, dispersed over both land and sea areas slightly affecting the coasts of Lombok Island (East of Bali; see Figure 1). By the spatial distribution of ash pixels within the plume, we can infer the different intensity of ash BTD signature. In particular, the latter was weaker where yellow/orange pixels were prevalent, i.e., especially in the distal plume region (which was presumably optically thinner) and on its borders. Figure 3b displays the RST ASH map of 28 November at 15:00 UTC (00:00 JST), providing information about plume shape and direction about two days after the previously analyzed satellite observation. As can be seen from the figure, although Mt. Agung was still emitting ash [48] the resulting plume was much smaller and less extended than that of Figure 3a, dispersing only over the Bali Island and moving in a SW direction. It is worth noting that RST ASH flagged a number of ash pixels even outside the magnified region on both panels of Figure 3 (e.g., see yellow pixels located at the NW corner of the maps). Those pixels, which were mostly associated to a low confidence level of detection for their location (they were not in spatial contiguity with the plume), presumably represented artefacts.
Information provided by Figure 3 is compatible with that reported by volcanological bulletins (e.g., continuous ash emission drifting WSW were observed during 28-29 November [48]), as well as with independent satellite-based observations of the same phenomenon corroborating the presence of ash over the ROI during the investigated period (see GOME-2 and TROPOMI absorbing aerosol index maps reported in [49,50]). However, to assess the ash-affected areas in a more accurate way, we performed a manual inspection of Himawari-8 imagery. In Figure 4 The plume, which may be well discriminated from both land (in red) and sea surfaces (in dark blue), as well as from meteorological clouds (in grey/white), had the same shape and direction of that flagged by RSTASH (see yellow polygon magnified at the top right side of the image). Nevertheless, it was more dispersed over Bali and Lombok Islands (e.g., see brown pixels close to the eruptive center), indicating that the used algorithm was slightly less effective in detecting ash both in the proximal and distal region of the plume. This limitation, which was already observed in previous studies analyzing ash clouds from different volcanoes (e.g., Eyjafjallajökull; Mt. Shinmoedake), may be ascribed to plume features (e.g., high optical thickness) and/or to observational conditions (e.g., low temperature contrast between the cloud-top and the underlying land surface) (e.g., [44,45]). Those factors presumably affected also the RSTASH map of Figure 3b, although cloud coverage had the major impact on ash detection results masking a significant portion of the plume (as for analyzed satellite data of 27 November, which are not shown here). Regarding ash pixels flagged by RSTASH outside the region marked in black, Figure 4 shows that those features mostly occurred in correspondence to The plume, which may be well discriminated from both land (in red) and sea surfaces (in dark blue), as well as from meteorological clouds (in grey/white), had the same shape and direction of that flagged by RST ASH (see yellow polygon magnified at the top right side of the image). Nevertheless, it was more dispersed over Bali and Lombok Islands (e.g., see brown pixels close to the eruptive center), indicating that the used algorithm was slightly less effective in detecting ash both in the proximal and distal region of the plume. This limitation, which was already observed in previous studies analyzing ash clouds from different volcanoes (e.g., Eyjafjallajökull; Mt. Shinmoedake), may be ascribed to plume features (e.g., high optical thickness) and/or to observational conditions (e.g., low temperature contrast between the cloud-top and the underlying land surface) (e.g., [44,45]).
Those factors presumably affected also the RST ASH map of Figure 3b, although cloud coverage had the major impact on ash detection results masking a significant portion of the plume (as for analyzed satellite data of 27 November, which are not shown here). Regarding ash pixels flagged by RST ASH outside the region marked in black, Figure 4 shows that those features mostly occurred in correspondence to some cloudy areas, confirming the generation of artefacts. The latter affected only a small portion of the entire analyzed satellite scene and did not prevent them from the correct identification of the ash plume.
Remote Sens. 2018, 10, x FOR PEER REVIEW 7 of 16 some cloudy areas, confirming the generation of artefacts. The latter affected only a small portion of the entire analyzed satellite scene and did not prevent them from the correct identification of the ash plume. To assess the RSTASH potential in monitoring the Agung ash cloud in real time, by exploiting the high temporal sampling of Himawari-8 data, we show in Figure 5 a time sequence (at 10 min) of ash maps referring to period 25 November at 21:10 UTC-26 November at 00:50 UTC (26 November at 06:10-09:50 JST). The figure shows the progressive displacement of airborne ash from the eruptive center (in purple) towards the Lombok Island, revealing also changes in both plume shape and direction. In detail, during the initial monitoring phase (i.e., on 25 November at 21:10-21:40 UTC) the ash plume, moving towards east and covering an area estimated around 100-200 km 2 , appeared very small because of aforementioned detection issues (see few ash pixels flagged over the Bali Island). Between 21:50-22:00 UTC, the plume direction changed from E to SE. In the following time intervals (i.e., since 22:10 UTC), the number of ash pixels increased revealing the continuity of ash emissions (e.g., see pixels flagged close to the eruptive center). Specifically, the strong increment of red pixels (i.e., those associated to low values of the   ( , , ) index) seems to indicate that a possible change in the ash plume features occurred (e.g., in terms of ash content and/or particle size). Afterwards, the monitored ash phenomenon did not show any significant variation. During the last monitoring phase (i.e., 26 November at 00:10 UTC-00:50 UTC), i.e., when the plume shape slightly changed, the number of ash pixels flagged over the Bali Island showed a small reduction, probably because of overlying meteorological clouds. Nonetheless, we recorded a general increase of ash coverage whose maximum value, around 900 km 2 , was retrieved from Himawari-8 data of 26 November at 00:50 UTC (when we can speculate that the amount of emitted ash was larger). To assess the RST ASH potential in monitoring the Agung ash cloud in real time, by exploiting the high temporal sampling of Himawari-8 data, we show in Figure 5 a time sequence (at 10 min) of ash maps referring to period 25 November at 21:10 UTC-26 November at 00:50 UTC (26 November at 06:10-09:50 JST). The figure shows the progressive displacement of airborne ash from the eruptive center (in purple) towards the Lombok Island, revealing also changes in both plume shape and direction. In detail, during the initial monitoring phase (i.e., on 25 November at 21:10-21:40 UTC) the ash plume, moving towards east and covering an area estimated around 100-200 km 2 , appeared very small because of aforementioned detection issues (see few ash pixels flagged over the Bali Island). Between 21:50-22:00 UTC, the plume direction changed from E to SE. In the following time intervals (i.e., since 22:10 UTC), the number of ash pixels increased revealing the continuity of ash emissions (e.g., see pixels flagged close to the eruptive center). Specifically, the strong increment of red pixels (i.e., those associated to low values of the ⊗ ∆TIR (x, y, t) index) seems to indicate that a possible change in the ash plume features occurred (e.g., in terms of ash content and/or particle size). Afterwards, the monitored ash phenomenon did not show any significant variation. During the last monitoring phase (i.e., 26 November at 00:10 UTC-00:50 UTC), i.e., when the plume shape slightly changed, the number of ash pixels flagged over the Bali Island showed a small reduction, probably because of overlying meteorological clouds. Nonetheless, we recorded a general increase of ash coverage whose maximum value, around 900 km 2 , was retrieved from Himawari-8 data of 26 November at 00:50 UTC (when we can speculate that the amount of emitted ash was larger). Therefore, even if RSTASH was less effective in detecting ash when eruption was probably at the early stage and the plume was presumably nearly opaque (i.e., during the initial monitoring phase), Therefore, even if RST ASH was less effective in detecting ash when eruption was probably at the early stage and the plume was presumably nearly opaque (i.e., during the initial monitoring phase), it was capable of providing reliable information about the space-time evolution of the monitored ash phenomenon. Indeed, despite the slight increase of artefacts characterizing some panels in Figure 5 (e.g., see those of 26 November at 00:00-00:50 UTC), RST ASH showed similar performance regardless of analyzed time slots and of considered ∆t. It should be stressed, however, that when cloud coverage was particularly significant affecting a large portion of the ROI (e.g., during 28 November at 14:20-15:40 UTC), the ash plume was more difficult to be identified and tracked by satellite. Hence, meteorological clouds, partially or completely obscuring the underlying ash layers, may still represent a major issue for operational monitoring of ash phenomena from space even when satellite data at high temporal resolution are used.

Estimates of Ash Cloud-Top Height
The plume height is one of key parameters used by numerical models aiming at forecasting ash dispersion in the atmosphere (e.g., [51]). This parameter may be retrieved from satellite observations by means of different procedures (e.g., [52][53][54][55][56][57]). Among them, the cloud-top temperature method assumes that ash clouds have an emissivity close to unity and their top is in thermodynamic equilibrium with the ambient air (e.g., [52][53][54]). The ash cloud-top height may then be estimated by satellite, comparing the minimum plume temperature measured in the TIR band (generally at around 11 µm wavelength) with the atmospheric temperature profile of ash-affected areas (e.g., [58][59][60]). This approach, like other methods, is affected by some limitations (e.g., it is better applicable to opaque clouds, whereas in the presence of optically thinner clouds or ash clouds close to the tropopause, results are generally less reliable; e.g., [60,61]). We have retrieved the plume-top height by comparing the temperature of the coldest ash pixel (in the AHI band 13) flagged by RST ASH , on each panel of Figure 5, to air temperature profiles made available by the NCEP (National Centers for Environmental Prediction) Reanalysis dataset [62]. Those profiles are provided 4-times per day at 17 different pressures levels, with a spatial sampling of 2.5 × 2.5 degrees. Figure 6 displays the temporal trend of the plume-top height (with a temporal sampling of 10 min) during the period 25 November at 21:10 UTC-26 November at 00:50 UTC (i.e., the same of Figure 5).
Remote Sens. 2018, 10, x FOR PEER REVIEW 9 of 16 it was capable of providing reliable information about the space-time evolution of the monitored ash phenomenon. Indeed, despite the slight increase of artefacts characterizing some panels in Figure 5 (e.g., see those of 26 November at 00:00-00:50 UTC), RSTASH showed similar performance regardless of analyzed time slots and of considered ∆t. It should be stressed, however, that when cloud coverage was particularly significant affecting a large portion of the ROI (e.g., during 28 November at 14:20-15:40 UTC), the ash plume was more difficult to be identified and tracked by satellite. Hence, meteorological clouds, partially or completely obscuring the underlying ash layers, may still represent a major issue for operational monitoring of ash phenomena from space even when satellite data at high temporal resolution are used.

Estimates of Ash Cloud-Top Height
The plume height is one of key parameters used by numerical models aiming at forecasting ash dispersion in the atmosphere (e.g., [51]). This parameter may be retrieved from satellite observations by means of different procedures (e.g., [52][53][54][55][56][57]). Among them, the cloud-top temperature method assumes that ash clouds have an emissivity close to unity and their top is in thermodynamic equilibrium with the ambient air (e.g., [52][53][54]). The ash cloud-top height may then be estimated by satellite, comparing the minimum plume temperature measured in the TIR band (generally at around 11 µ m wavelength) with the atmospheric temperature profile of ash-affected areas (e.g., [58][59][60]). This approach, like other methods, is affected by some limitations (e.g., it is better applicable to opaque clouds, whereas in the presence of optically thinner clouds or ash clouds close to the tropopause, results are generally less reliable; e.g., [60,61]). We have retrieved the plume-top height by comparing the temperature of the coldest ash pixel (in the AHI band 13) flagged by RSTASH, on each panel of Figure 5, to air temperature profiles made available by the NCEP (National Centers for Environmental Prediction) Reanalysis dataset [62]. Those profiles are provided 4-times per day at 17 different pressures levels, with a spatial sampling of 2.5 × 2.5 degrees. Figure 6 displays the temporal trend of the plume-top height (with a temporal sampling of 10 min) during the period 25 November at 21:10 UTC-26 November at 00:50 UTC (i.e., the same of Figure 5). The green curve shows that the plume-top height fluctuated in the range of 4.8-9.3 km during the investigated period. In particular, the analyzed parameter showed small variations on 25 November until 23:10 UTC (when it was in the range 4.8-5.4 km), whereas since 23:40 UTC it progressively increased despite some fluctuations (also because of cloud coverage affecting ash detection results). A sudden increase of plume-top height (from 7.4 km to 9.1 km) was recorded on 26 November between 00:00-00:10 UTC, revealing an abrupt variation in the intensity of ash The green curve shows that the plume-top height fluctuated in the range of 4.8-9.3 km during the investigated period. In particular, the analyzed parameter showed small variations on 25 November until 23:10 UTC (when it was in the range 4.8-5.4 km), whereas since 23:40 UTC it progressively increased despite some fluctuations (also because of cloud coverage affecting ash detection results). A sudden increase of plume-top height (from 7.4 km to 9.1 km) was recorded on 26 November between 00:00-00:10 UTC, revealing an abrupt variation in the intensity of ash emissions. Afterwards, the analyzed parameter slightly decreased, increasing once again (up to the maximum value) at 00:50 UTC. Figure 6 displays also values (in red) of the plume height reported by Darwin VAAC, which were independently retrieved from Himawari-8 imagery and CVGHM (Center for Volcanology and Geological Hazard Mitigation) ground reports. According to the VA (Volcanic Ash) advisors from this monitoring center (whose monitored area includes Indonesia, Papua New Guinea and part of the Philippines [63]), the ash plume was at FL200 (~6.1 km) on 25 November at 21:00 UTC, increased to FL220 (~6.7 km) at 22:20 UTC and extended up to FL260 (~7.9 km) on 26 November at 00: 40 UTC [64]. Hence, although only a few independent estimates of plume height were available during the analyzed time interval, it seems that the ash plume was, in general, higher in altitude than indicated by the green curve in Figure 6. Considering that during the initial monitoring phase RST ASH was less effective in detecting ash, particularly close to the source of emissions (because of some limitations previously discussed in Section 4.1), it is realistic to suppose that the plume height was consequently underestimated (of about 28% in reference to data of 22 November at 22:20 UTC). On the other hand, estimates of this parameter became presumably more accurate when RST ASH performed better in detecting the proximal region of the ash cloud, as indicated by the good agreement with information provided by Darwin VAAC (see 26 November at 00:40 UTC).

Discussion
Results achieved investigating the ash plume emitted during the Mt. Agung eruptions of 25-28 November 2017 have shown that RST ASH may benefit from the improved features of the AHI sensor onboard the new Japanese geostationary satellites (including Himawari-9, which is currently on standby). Information provided by RST ASH maps (in terms of ash location, plume shape and direction) is in agreement with independent field and satellite-based observations. The usefulness of those maps is even more evident considering that the ash RGB product specifically designed for Himawari-8 (see Section 2) did not always enable the identification of airborne ash during the period of interest. An example is reported in Figure 7a ). The figure shows that volcanic ash, which should appear in pinkish color (see [25,27]), was difficult to be distinguished from the background (see region magnified at the top right side of the image). As can be seen from Figure 7b, modifying the Red component, i.e., analyzing the BT 11.2 − BT 10.4 spectral difference in place of BT 12.4 − BT 10.4 , the ash plume was more recognizable but at the expense of an increase of misleading features, which were mostly associated to meteorological clouds (e.g., see pink areas outside the region marked in black).
We have verified the RST ASH capacity in detecting airborne ash despite: (i) the absence of a highly populated dataset of satellite records and (ii) a different implementation than usual. In the first case, although we processed only three years of satellite records (i.e., 90 images for each analyzed time slot), the generated spectral reference fields allowed us to correctly run the algorithm. It is important to stress that RST ASH could be in principle used even in absence of a plurennial dataset of satellite records, by analyzing, for instance, data belonging to contiguous time slots for increasing statistics (as suggested by other works performed using the RST approach [65,66]). Regarding the algorithm implementation, the generation of the library of spectral reference fields may be time demanding, particularly when geostationary satellite imagery are analyzed because of the large amount of data to be processed. In this study, we have reduced the number of spectral reference fields to run RST ASH . This choice had a low impact on ash detection results, as discussed in Section 4.2. Figure 8, showing the ( 10.4 − 11.2 ) and ( 3.9 − 10.4 ) images in reference to three different t (i.e., 10 min; 30 min; 60 min), seems to corroborate the low dependency of spectral reference fields on short-time intervals. Indeed, Figure 8a shows that the spectral difference 10.4 − 11.2 tends to fluctuate less (despite the presence of clouds) regardless of the analyzed t (as indicated by the mean value around zero and standard deviation lower than one of the histograms associated to top panel images). Figure 8b shows that the 3.9 − 10.4 signal fluctuates in a more significant way in daylight conditions owing to solar reflection effects affecting the MIR signal (see higher mean and This choice had a low impact on ash detection results, as discussed in Section 4.2. Figure 8, showing the ∆(BT 10.4 − BT 11.2 ) and ∆(BT 3.9 − BT 10.4 ) images in reference to three different ∆t (i.e., 10 min; 30 min; 60 min), seems to corroborate the low dependency of spectral reference fields on short-time intervals. Indeed, Figure 8a shows that the spectral difference BT 10.4 − BT 11.2 tends to fluctuate less (despite the presence of clouds) regardless of the analyzed ∆t (as indicated by the mean value around zero and standard deviation lower than one of the histograms associated to top panel images). Figure 8b shows that the BT 3.9 − BT 10.4 signal fluctuates in a more significant way in daylight conditions owing to solar reflection effects affecting the MIR signal (see higher mean and standard deviation values of histograms associated to bottom panel images). Nonetheless, by comparing the BT 10.4 − BT 11.2 and BT 3.9 − BT 10.4 signals to relative spectral reference fields of t ± 30 min (i.e., the central panels in Figure 8) rather than t, only a slight reduction of RST ASH performance is expected, thanks to the joint usage of the two local variation indices described in Section 2.
Remote Sens. 2018, 10, x FOR PEER REVIEW 12 of 16 standard deviation values of histograms associated to bottom panel images). Nonetheless, by comparing the 10.4 − 11.2 and 3.9 − 10.4 signals to relative spectral reference fields of t ± 30 min (i.e., the central panels in Figure 8) rather than t, only a slight reduction of RSTASH performance is expected, thanks to the joint usage of the two local variation indices described in Section 2. The maximum time interval ∆t = 30 min could then be selected to perform a full algorithm implementation on Himawari-8 data, allowing us to analyze 48 instead of 144 time slots. This implementation scheme, which might be applied even to data provided by the next MTG-FCI The maximum time interval ∆t = 30 min could then be selected to perform a full algorithm implementation on Himawari-8 data, allowing us to analyze 48 instead of 144 time slots. This implementation scheme, which might be applied even to data provided by the next MTG-FCI (Meteosat Third Generation-Flexible Combined Imager) geostationary satellites (providing additional spectral channels and a better spatial, temporal and radiometric resolution than MSG-SEVIRI) could allow us to rapidly develop an automated RST ASH elaboration chain to detect ash clouds in real time. The traditional RST ASH implementation scheme is, however, still the suggested one, particularly when an extensive data archive and adequate computational resources are available (the algorithm runs fast once the library of spectral reference fields is generated).
Finally, the short-term temporal fluctuations of the ash plume-top height shown in Figure 6 seem to indicate that estimates of this parameter were, in general, more accurate when RST ASH performed better in detecting ash in the vicinity of the eruptive center (although also intrinsic limitations of the cloud-top temperature method probably affected the achieved results). The study confirms the main issues of RST ASH , including the generation of residual artefacts occurring in correspondence to some meteorological clouds (an optimized algorithm configuration tested with success in a previous study could be used for minimizing those features on daytime satellite scenes [46]). Those artefacts that were mostly associated to pixels with a medium-low confidence level of detection on the one side enable a better mapping of ash coverage, but on the other side may lead to an increase of false positives (this aspect should be taken into account if retrieval analyses, starting from RST ASH detections, are performed).

Conclusions
This paper shows that the RST ASH performance coupled to high temporal resolution of Himawari-8 data may lead to an effective identification and tracking of ash clouds over East Asia and Western Pacific region in spite of some limitations. Some investigations are currently in progress to assess if the used algorithm may benefit from the analysis of BT 8.7 − BT 10.4 signal in the space-time domain. A specific local variation index based on this spectral difference should guarantee a further reduction of artefacts, in both daylight and nighttime conditions, if integrated within the RST ASH detection scheme. This way, RST ASH could provide an even better contribution for supporting activities devoted to mitigate the impact of strong explosive eruptions on air traffic in the framework of operational contexts.
Author Contributions: F.M., N.P. and V.T conceived the research work. F.M wrote the majority of the paper. A.F. implemented the RST ASH algorithm on Himawari-8 data and performed the tests. N.P. and V.T. contributed in writing the paper.
Funding: This research was funded by European Union's Horizon 2020 research programme for Societal challenges-smart, green and integrated transport, Research Project EUNADICS-AV, grant number 723986.