Assessing Performance of the RST VOLC Multi-Temporal Algorithm in Detecting Subtle Hot Spots at Oldoinyo Lengai (Tanzania, Africa) for Comparison with MODLEN

: The identiﬁcation of subtle thermal anomalies (i.e., of low-temperature and/or spatial extent) at volcanoes by satellite is of great interest for scientists, especially because minor changes in surface temperature might reveal an unrest phase or impending activity. A good test case for assessing the sensitivity level of satellite-based methods is to study the thermal activity of Oldoinyo Lengai (OL) (Africa, Tanzania), which is the only volcano on Earth emitting natrocarbonatite lavas at a lower temperature (i.e., in the range 500–600 ◦ C) than usual magmatic surfaces. In this work, we assess the potential of the RST VOLC multi-temporal algorithm in detecting subtle hot spots at OL for comparison with MODLEN: A thermal anomaly detection method tailored to OL local conditions, by using Moderate Resolution Imaging Spectroradiometer (MODIS) data. Our results investigating the eruptive events of 2000–2008 using RST VOLC reveal the occurrence of several undocumented thermal activities of OL, and may successfully integrate MODLEN observations. In spite of some known limitations strongly affecting the identiﬁcation of volcanic thermal anomalies from space (e.g., cloud cover; occurrence of short-lived events), this work demonstrates that RST VOLC may provide a very important contribution for monitoring the OL, identifying subtle hot spots showing values of the radiant ﬂux even around 1 MW. hectares or km 2 at other active volcanoes. Due to the limited area and temperature of active lava ﬂows, the OL thermal activity is generally difﬁcult to investigate by satellite. A speciﬁc adaption of MODVOLC, named MODLEN, has been designed, tailored and tested with good results for better identifying subtle hot spots at OL [30–33]. This is not routinely monitored by ground-based systems, as is similar with other African volcanoes [34].

The synoptic view of satellite systems along with the frequent observations enable active volcanoes monitoring both in developing countries, where economic resources are generally insufficient to install ground-based surveillance systems [9], as well as in areas well monitored by geophysical devices (e.g., Reference [15]). Among the space-based sensors, suited to monitor active volcanoes, several studies have been performed using the AVHRR (Advanced Very High Resolution Radiometer), aboard NOAA (National Oceanic and Atmospheric Administration) and METOP (Meteorological Operational (Meteorological Operational Satellites) satellites (e.g., [16][17][18][19][20]). MODIS (Moderate Resolution Imaging Spectroradiometer), aboard NASA (National Aeronautics and Space Administration) Earth Science satellite missions Terra and Aqua., due to its spectral features, and to a good trade-off between temporal (4 passages per day) and spatial (1 km) resolution, has further extended detection capabilities of active lava flows, enabling also their quantitative characterization (e.g., [20][21][22][23][24][25][26]). MODIS offers in fact some spectral channels in the MIR (Medium InfraRed) band sensitive to hot surfaces, with the channel 21 (3.929-3.989 µm) having a better dynamic range than the other space-based IR sensors (it saturates at nearly 500 K in comparison with ~330-340 K of traditional MIR systems).
Among the satellite-based algorithms developed to detect volcanic hot spots using MODIS data, MODVOLC [20] (operating since 2000 at NASA Distributed Active Archive Center (DAAC)) performs the near-real time monitoring of active volcanoes at a global scale, e.g., References [20][21][22][23]. MIROVA (Middle Infrared Observation of Volcanic Activity) has been operational since 2013, providing information about thermal anomalies and Volcanic Radiative Power (VRP), at several volcanoes within 1 to 4 h of each satellite overpass [25,26]. RSTVOLC is a multi-temporal method that was tested with success in different geographic areas [27,28], revealing the occurrence of possible thermal precursory signals of some explosive eruptions, e.g., Reference [24].
In this work, we investigate the thermal activity of Oldoinyo Lengai (Africa-Gregory Rift Valley; see Figure 1) from space. This remote volcano is the only volcano on Earth emitting natrocarbonatite lava that shows a very low level of viscosity (10 −1 to 10 2 Pa s [29]), and a relatively low temperature (ranging from 495-590 °C), compared to other types of lavas [30,31]. The eruption rates at Oldoinyo Lengai (OL) are extremely low; therefore, lava covers only a few tens of m 2 compared to several hectares or km 2 at other active volcanoes. Due to the limited area and temperature of active lava flows, the OL thermal activity is generally difficult to investigate by satellite. A specific adaption of MODVOLC, named MODLEN, has been designed, tailored and tested with good results for better identifying subtle hot spots at OL [30][31][32][33]. This is not routinely monitored by ground-based systems, as is similar with other African volcanoes [34]. Here we analyze the OL eruptive activity between 2000-2008, through using RSTVOLC, assessing the performance of this largely accepted method for comparison with MODLEN, which may be considered as the benchmark for this volcanic area. Here we analyze the OL eruptive activity between 2000-2008, through using RST VOLC , assessing the performance of this largely accepted method for comparison with MODLEN, which may be considered as the benchmark for this volcanic area.
During the selected period, the OL showed different eruption styles. Moreover, many field observations were available for validation (because of changes in the landscape and crater morphology, the access to summit crater has become more difficult after 2008). Bulletins of the Global Volcanism Program (GVP) and other independent field reports, made available by Frederick Belton on his website [35], are used here to validate thermal anomalies identified by satellite. In addition, a manual inspection of infrared MODIS records [36] is performed, and available high spatial resolution infrared satellite data are also used for assessing results.
The work focuses on the identification of subtle thermal anomalies (i.e., those of low temperature and/or small size). This is of a great interest for understanding volcanic processes (e.g., small explosive eruptions) as well as for other remote sensing applications (e.g., small fires characterization [36]).

Oldoinyo Lengai Eruptive Activity
Two types of eruptive activity associated with contrasted magma composition characterize OL. In 1917In , 1940In -1941In , 1967, and more recently in 2007-2008, vulcanian to sub-plinian explosive eruptions occurred over weeks to months, generating ash plumes as high as 11 km spreading ash over the entire region [30,32,[37][38][39]. These eruptions are associated with nephenelite magma compositions [39]. Explosive eruptions are generally followed by quiescent periods. Renewal of eruptive activity is characterized by the effusion of low viscosity natrocarbonatite lava flows gradually filling up the crater created by the explosive eruption, and overflowing on the volcano flanks. This was observed before the 1917 and 1966 eruptions and from January 1983 onwards, until the 2007-2008 eruption events. Evidence of lava effusion was observed again by late 2008, after the end of the explosive eruption. Our analysis focuses specifically on the phase of natrocarbonatite effusion from 2000 to 2008. From 2000 to 2004, the crater was filled with lava which spilled over three points on the crater rim descending hundreds of meters down the outer flanks [40,41]. Eruptive activity was fed from several hornitos, scattered across the 400 m wide platform filling the crater. In July 2004, eruptions became more explosive producing small ash clouds [42][43][44]. During late March and early April 2006, a vigorous effusive eruption occurred. A lava flow with exceptionally large volume (9 × 10 5 m 3 ; [31]) was emitted in two phases and was emplaced along the western flank, reaching a length of 3 km (see [31]). After approximately one year of quiescence, a new lava emission occurred in July and August 2007, from several vents opening within the crater. This effusive activity was slightly more intense than that observed over the last 25 years [32,37,39]. On 4 September 2007, intense explosive activity started with repetitive emission of ash plumes, varying from several hundred meters to 11 km elevation [32,37]. This episodic explosive activity culminated in February-March 2008, with the generation of small pyroclastic density currents along the flank of the volcano and the formation of a 200 m deep pit crater. The analysis of ash samples from 2007-2008 indicated that the eruptive activity produced a hybrid magma generated by the interaction between natrocarbonatite and nephelinite [39,45].

Satellite Methods
As before mentioned, the aim of this paper is to assess the RST VOLC potential in detecting subtle hot spots at OL using as a reference MODLEN, which was specifically developed for monitoring this African volcano. Although other MODIS-based methods (e.g., MIROVA) could be analyzed for assessing differences in detecting volcanic thermal anomalies at Lengai during the period of interest, an inter-comparison among several hot spot detection methods it out of the scope of this paper. In the following paragraphs, the rationale of MODLEN and RST VOLC algorithms is then described.

MODLEN
MODLEN is an adaptation of the global MODVOLC algorithm, developed by Wright and Pilger [21], to monitor volcanic hot spots globally. The algorithm is based on the derivation of the Normalized Thermal Index (NTI), which is defined as: in which L MIR (x,y) and L TIR (x,y) are the radiance values of MIR and TIR (Thermal Infrared) signals measured in the MODIS channel 22 (3.929-3.989 µm), or channel 21 when channel 22 is saturated, and 32 (11.770-12.270 µm), respectively, for a given pixel. MODVOLC works with a single threshold value for all volcanoes in the world: All pixels returning an NTI value above −0.80 are considered as hot spots [21]. This threshold, although guarantees a reliable identification of thermal anomalies at a global scale, was not found to be sensitive to OL eruptive activity [28]. To overcome this limitation, MODLEN introduces a regionally valid threshold, checking for all the pixels of the studied area, and if the NTI values are above −0.83, in this case automatically flagging the pixel as a hot spot. For pixels with an NTI value between −0.83 and −0.88, a spatial filter is applied to compare the NTI values of the studied pixel with the average of its eight surrounding neighbors (NTI neigh ), according to: Because NTI values are negative, and a sub-pixel thermal anomaly will generate a higher-than-average NTI value. The spatial derivative will be negative for hot spots. Empirical analysis showed that anomalously hot pixels return values of the spatial filter lower than −0.02. MODLEN. Therefore, flags all pixels with a spatial derivative smaller than −0.02 and a NTI greater than −0.88 as hot spots. Since the MODLEN algorithm returns low NTI values for cloud-affected pixels, no cloud mask is applied. The use of this spatial derivative parameter turns the point operation of the MODVOLC algorithm into a contextual approach, where the presence of an anomaly at one pixel is evaluated taking into account the radiance measured at the surrounding pixels. Eventually, in order to facilitate the interpretation of detected hot spots, the MODLEN algorithm automatically computes the distance of any identified hot spot to OL crater. This disregards hot spots generated by bush fires (generally because of lava overflowing) at the base of the volcano. The NTI value of the pixel located closest to Oldoinyo Lengai active crater is also recorded for each analyzed scene, even if it is not flagged as a hot spot in order to assess the effect of cloud coverage (with typical NTI value lower than −0.92). Hot spots detected by MODLEN have been shown to generally match with eruptive activity, witnessed at OL [28]. When observations were available, they proved to be essential data to constrain the changes and variation in eruptive activity from 2000 to 2008 [31,32,46].

RST VOLC
RST VOLC is an optimized configuration of the RST (Robust Satellite Techniques) multi-temporal approach [47], which identifies perturbing events by: (i) Analyzing multiyear time series of cloud-free homogeneous (same calendar month, same channel/s, same acquisition time) satellite records; and (ii) searching for anomalies by means of a specific change detection step.
In more detail, RST VOLC uses two local variation indices in combination to identify volcanic thermal anomalies [27]: Remote Sens. 2018, 10, 1177 5 of 17 In Equation (3), BT MIR (x, y, t) is the brightness temperature (BT), measured in the MODIS channel 22 at time t for each pixel (x, y) of the analysed satellite scene. While µ MIR (x, y) and ⊗ MIR (x, y), respectively, represent the temporal mean and the standard deviation. As for MODLEN, channel 21 is used in place of channel 22 when the latter is saturated.
To minimize those spurious effects, the ⊗ MIR−TIR (x, y, t) index, analyzing also the signal measured in the TIR band at around 11 µm wavelength (i.e., BT TIR (x, y, t)), is used in combination with the previous one. The ⊗ MIR−TIR (x, y, t) index is described in Equation (4), where ∆T(x, y, t) = BT MIR (x, y, t) − BT TIR (x, y, t), while µ ∆T (x, y) and σ ∆T (x, y) stand for the relative temporal mean and standard deviation.
Along with previous studies, we have used the OCA (One Channel Algorithm [51]) method to filter out meteorological clouds, and the iterative 2σ-clipping filter to remove signal outliers possibly contaminating the time series, before generating the library of spectral reference fields (i.e., temporal mean and standard deviation images) [27,28]. OCA is an RST-based cloud-detection scheme that identifies cloudy radiances based on signal divergence from unperturbed "clear sky" conditions. In particular, only the TIR signal is analyzed in nighttime, which considers the different spectral behavior shown by those features in comparison with the cloud-free background. The MODIS channel 32 (11.770-12.270 µm) is usually used for this purpose. The OCA method performs well in different geographic areas and under different illumination conditions, in spite of some limitations widely discussed in previous works [52].

Data and Algorithm Implementation
MODIS is a sensor flying aboard the Terra (morning coverage) and Aqua satellites (afternoon coverage), providing data in 36 spectral channels. The infrared channels have a spatial resolution of 1 km at nadir view. Channels 21 and 22, which are centered in the MIR spectral region, and channels 31 (10.780-11.280 µm) or 32, located in the TIR, are commonly used to detect volcanic hot spots (e.g., [21]). Specifically, the channel 21 is generally used when the channel 22 saturates. In this work, to assess RST VOLC performance at OL, we have processed nine years of nighttime MODIS records acquired from 2000 to 2008, provided by NASA (LAADS-Level 1 and Atmosphere Archive and Distribution System). Table 1 details the number of analyzed Aqua (2171) and Terra (2593) orbits (4764 images in total). We have used MODIS data, as listed in Table 1, to generate the spectral reference fields (on a monthly temporal scale) required for the computation of RST VOLC indices. Two different time slots (of 19:00 UTC ± 2 h for Terra and 01:00 UTC ± 2 h for Aqua) have been analyzed for this purpose. Thus, to implement RST VOLC , running on re-projected Level 1b MODIS imagery (in LAT/LONG WGS 84 projection) with a 1 km resampled ground resolution cell, we have generated a library of 48 spectral reference fields (i.e., one temporal mean and one standard deviation image for each time slot and month of interest and for each sensor).

1.
We looked for the temporal agreement between flagged hot spots and documented eruptive activities of OL (e.g., [40][41][42][43][44]). The GVP bulletins usually provide information about the OL based on visual observations performed by visitors, single researchers and/or scientific teams. Since this information is generally less detailed than that reported by Frederick Belton (who carried out several field observations at OL during the period of interest) on his website [35]. We used both independent sources to assess thermal anomalies flagged by satellite.

2.
Field observations are infrequent and intermittent at OL. Therefore, it is realistic to suppose that a number of thermal activities occurring during the period of interest were unreported by both validation sources [30]. In order to take this limitation into account, and also considering the lava cooling process, we considered that thermal anomalies flagged within ±2 days from documented eruptions were associated to real (although undocumented) thermal activities of OL.

3.
High spatial resolution satellite data provided by sensors as ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) and ALI (Advanced Land Imaging) can also be Remote Sens. 2018, 10, 1177 7 of 17 used for assessing the presence of volcanic hot spots at OL. The low temporal resolution of those data (e.g., 16 days for both above-mentioned sensors), which is further reduced in presence of clouds, did not allow us for a systematic validation analysis of the detected thermal anomalies. As an example, by analyzing ASTER data made available from AVA (Aster Volcano Archive; https://ava.jpl.nasa.gov/granule.php?f=20040805200215) of the JPL (Jet Propulsion Laboratory), we found that only 13 granules independently returned hot spots for the OL during the period of interest, confirming limitations mentioned above (see Table 3). Considering only the ASTER hot spots at an average distance less than 1 km from the OL summit, eight of them indicate the occurrence of volcanic activity and six were temporally coincident with thermal anomalies identified by MODLEN and/or RSTVOLC.

4.
To assess if thermal anomalies that were not corroborated by the three above-mentioned validation steps represented artefacts or were rather associated to undocumented thermal activities of OL, we performed a manual inspection of infrared MODIS records, similar to previous independent studies [36]. This analysis, which was carried out by assessing the location and features (in terms of brightness temperature difference with the background) of detected hot spots, is similar to that independently performed in a previous literature study. This is where authors considered anomalous those pixels whose integrated temperature was elevated by 5 K above its neighbor (non-anomalous) ones [36,53]. In Figure Table 4 summarizes results of the multi-step analysis described above indicating that 92.2% and 97.7%, respectively, of MODLEN and RST VOLC detections were associated to real (i.e., both documented and undocumented) thermal activities of OL. It should be stressed that a correlation analysis between the two analyzed hot spot detection series led to an adjusted R 2 = 0.87 and to a p-value < 0.0001, revealing the similar behavior of MODLEN and RST VOLC algorithms during the period of interest.   Table 4 summarizes results of the multi-step analysis described above indicating that 92.2% and 97.7%, respectively, of MODLEN and RSTVOLC detections were associated to real (i.e., both documented and undocumented) thermal activities of OL. It should be stressed that a correlation analysis between the two analyzed hot spot detection series led to an adjusted = 0.87 and to a -value < 0.0001, revealing the similar behavior of MODLEN and RSTVOLC algorithms during the period of interest.  After evaluating performance of RST VOLC in detecting volcanic hot spots for comparison with MODLEN (e.g., see false positive rates detailed in Table 4), we investigated differences in monitoring changes of thermal volcanic activity at OL. Figure 3 summarizes the results of this analysis, displaying the temporal trend of thermal anomaly spatial extent (in terms of hot spot pixel number) retrieved during the period of interest by matching the geographic coordinates of anomalous pixels, flagged by RST VOLC on original Level 1b MODIS data used by MODLEN. As can be seen from the figure, according to both RST VOLC (red bars) and MODLEN (grey bars) the most intense eruptive events of OL were those of March-April 2006 (when an unusually large effusive eruption of natrocarbonatite lava occurred) and July-September 2007 (when another significant eruptive event took place at the monitored volcano). It is worth noting that after the identification of an undocumented eruption on 20-21 June 2006, the two methods did not flag any thermal anomaly before June 2007, in agreement with the information provided by field reports [54]. After evaluating performance of RSTVOLC in detecting volcanic hot spots for comparison with MODLEN (e.g., see false positive rates detailed in Table 4), we investigated differences in monitoring changes of thermal volcanic activity at OL. Figure 3 summarizes the results of this analysis, displaying the temporal trend of thermal anomaly spatial extent (in terms of hot spot pixel number) retrieved during the period of interest by matching the geographic coordinates of anomalous pixels, flagged by RSTVOLC on original Level 1b MODIS data used by MODLEN. As can be seen from the figure, according to both RSTVOLC (red bars) and MODLEN (grey bars) the most intense eruptive events of OL were those of March-April 2006 (when an unusually large effusive eruption of natrocarbonatite lava occurred) and July-September 2007 (when another significant eruptive event took place at the monitored volcano). It is worth noting that after the identification of an undocumented eruption on 20-21 June 2006, the two methods did not flag any thermal anomaly before June 2007, in agreement with the information provided by field reports [54].  Figure 4 displays the temporal trend of the radiant flux ( ) retrieved from RSTVOLC detections performed on MODIS data of July-early September 2007 (similar results are expected from MODLEN outputs). We estimated the radiant flux using the formulation reported in Reference [55] (amending that proposed by Kaufman et al. [56] to retrieve the rate of emitted energy of active fires):

Quantifying Thermal Emissions at OL
In Equation (5), and respectively represent the MIR brightness temperatures of thermal anomaly and of background (the temporal mean value has been considered here); is the MODIS pixel area (1 km 2 for the resampled MODIS data); and is the number of hot spot pixels.  Figure 4 displays the temporal trend of the radiant flux (Q rad ) retrieved from RST VOLC detections performed on MODIS data of July-early September 2007 (similar results are expected from MODLEN outputs). We estimated the radiant flux using the formulation reported in Reference [55] (amending that proposed by Kaufman et al. [56] to retrieve the rate of emitted energy of active fires):

Quantifying Thermal Emissions at OL
In Equation (5), T h and T b respectively represent the MIR brightness temperatures of thermal anomaly and of background (the temporal mean value has been considered here); A pix is the MODIS pixel area (1 km 2 for the resampled MODIS data); and n is the number of hot spot pixels. Figure 4 shows that the radiant flux ranged from 1.2 MW (on MODIS data of 4 July at 23:05 UTC, when a single hot spot pixel was detected by RST VOLC ) to about 775 MW (on MODIS data of 1 September at 19:50 UTC when both algorithms flagged the same number of anomalous pixels). In more detail, the radiant flux was lower than 50 MW during 1-7 July (when some thermal anomalies were identified also by MODVOLC [46]). During 23 July-20 August 2007, RST VOLC detected volcanic hot spots in a more discontinuous way; the estimated radiant flux was mostly lower than 10 MW. According to field observations, an inner crater activity occurred during 22-23 July, small lava flows were emitted during 3-5 August and a mild-strombolian activity took place during 15-16 August, indirectly corroborating the low level of radiant flux in Figure 4. The latter also shows that the analyzed parameter increased in magnitude during 22-25 August because of a documented lava effusion, while it decreased in the following two days (when Q rad was retrieved around 7 MW on MODIS data of 27 August at 22:25 UTC).  Afterwards (i.e., during 28-30 August), a new increase in the intensity of thermal emissions was recorded, indicating the possible occurrence of a lava overflow or the presence of an active lava lake within the crater, in agreement with that reported in a previous study [34]. The RGB product from EO1-ALI imagery of 29 August 2007, shown in Figure 5, further confirms this hypothesis.
Indeed, the false color composite image shows that some pixels appearing red, because of strong Short Wave InfraRed (SWIR) signal, affected the NE side of OL crater corroborating the occurrence of a lava overflow. Three days later, once again increased reaching about 250 MW (comparable values of this parameter were retrieved on both Aqua and Terra-MODIS scenes), although the radiant flux reached its peak on 1 September, when the lava flowed down the volcano flanks [54]. Since the Afterwards (i.e., during 28-30 August), a new increase in the intensity of thermal emissions was recorded, indicating the possible occurrence of a lava overflow or the presence of an active lava lake within the crater, in agreement with that reported in a previous study [34]. The RGB product from EO1-ALI imagery of 29 August 2007, shown in Figure 5, further confirms this hypothesis.
Indeed, the false color composite image shows that some pixels appearing red, because of strong Short Wave InfraRed (SWIR) signal, affected the NE side of OL crater corroborating the occurrence of a lava overflow. Three days later, Q rad once again increased reaching about 250 MW (comparable values of this parameter were retrieved on both Aqua and Terra-MODIS scenes), although the radiant flux reached its peak on 1 September, when the lava flowed down the volcano flanks [54]. Since the voluminous lava overflow triggered some bush fires [34], the maximum Q rad value around 775 MW was presumably overestimated. Nonetheless, this value of the radiant flux appears not particularly high, especially if compared to estimates of same parameter performed in other volcanic areas in presence of active lava flows (e.g., [28]). Finally, Figure 4 shows that, after the above-mentioned lava effusion, intensity of volcanic thermal emissions significantly decreased, as indicated by the lower Q rad value, around 82 MW, retrieved from MODIS data of 3 September at 19:35 UTC. This analysis demonstrates that quantitative information retrieved from satellite observations may be helpful in reconstructing the dynamics of eruptive events.

Discussion
Results shown in previous section demonstrate that RSTVOLC was capable of detecting both natrocarbonatite lava events and minor thermal activities of OL (leading also to values of radiant flux around 1-10 MW generally associated to weak strombolian eruptions; e.g., [57]) occurring during the period of interest. In detail, analyzing the field reports, including 227 days of low-moderate thermal activity (e.g., gas emissions, weak strombolian explosions; lava within the crater floor) and 21 days of high-level eruptions (e.g., lava effusions/fountains), RSTVOLC identified about 13.7% of documented eruptive days, performing slightly better than MODLEN whose percentage was around 12.9%.
It is worth mentioning that those percentages were achieved assessing results based on the multi-step validation analysis detailed in Section 5.1 and then considering as real also hot spots flagged within the temporal range of ±2 days from documented eruptions. It should be remarked that such a criterion takes into account the limit of field observations that are infrequent and intermittent at OL, possibly missing the start/end time of eruptive activity. This hypothesis has been confirmed here: For example, by the analysis of thermal anomalies shown in Table 3. Indeed, of the eight thermal anomalies independently revealed by ASTER data at OL, five occurred within the selected ±2 days from documented eruptions.

Discussion
Results shown in previous section demonstrate that RST VOLC was capable of detecting both natrocarbonatite lava events and minor thermal activities of OL (leading also to values of radiant flux around 1-10 MW generally associated to weak strombolian eruptions; e.g., [57]) occurring during the period of interest. In detail, analyzing the field reports, including 227 days of low-moderate thermal activity (e.g., gas emissions, weak strombolian explosions; lava within the crater floor) and 21 days of high-level eruptions (e.g., lava effusions/fountains), RST VOLC identified about 13.7% of documented eruptive days, performing slightly better than MODLEN whose percentage was around 12.9%.
It is worth mentioning that those percentages were achieved assessing results based on the multi-step validation analysis detailed in Section 5.1 and then considering as real also hot spots flagged within the temporal range of ±2 days from documented eruptions. It should be remarked that such a criterion takes into account the limit of field observations that are infrequent and intermittent at OL, possibly missing the start/end time of eruptive activity. This hypothesis has been confirmed here: For example, by the analysis of thermal anomalies shown in Table 3. Indeed, of the eight thermal anomalies independently revealed by ASTER data at OL, five occurred within the selected ±2 days from documented eruptions.
As a further demonstration of the consistency of this criterion, Figure 6 displays the EO1-ALI RGB product at 30 m spatial resolution of 10 March 2008, when both MODLEN and RST VOLC flagged a thermal anomaly over the OL. The figure shows the presence of some highly radiant pixels in the SWIR band well located over the crater area (see red pixels), confirming that, one day before the occurrence of a documented eruptive activity (of 11 March 2008), a hot magmatic surface already affected the monitored volcanic area. It should be pointed out that the two algorithms provided comparable information about thermal anomalies occurring within the temporal range of ±2 days, which were in the order of about 28% for RST VOLC and 24% for MODLEN, as a further demonstration of their similar performance.
Regarding the capability of both methods in monitoring changes of OL thermal activity also investigated in this study (see Figure 3), main differences between them are ascribable to their intrinsic features. As an example, MODLEN considered the July-August 2007 eruption as less intense than that indicated by RST VOLC mainly because of the used spatial filter, favoring the identification of the brightest pixels in a scene within 2 km of Lengai summit area and then causing an underestimation of thermal features. On the other hand, RST VOLC possibly overestimated some volcanic hot spots owing to bush fires developing along the volcano flanks (the algorithm does not include any spatial-length filter to remove those features unlike MODLEN) and/or because of pixel duplication effects occurring during data resampling.
Remote Sens. 2018, 10, x FOR PEER REVIEW 12 of 16 intrinsic features. As an example, MODLEN considered the July-August 2007 eruption as less intense than that indicated by RSTVOLC mainly because of the used spatial filter, favoring the identification of the brightest pixels in a scene within 2 km of Lengai summit area and then causing an underestimation of thermal features. On the other hand, RSTVOLC possibly overestimated some volcanic hot spots owing to bush fires developing along the volcano flanks (the algorithm does not include any spatial-length filter to remove those features unlike MODLEN) and/or because of pixel duplication effects occurring during data resampling. Despite the differences mentioned above, the similar behavior of RSTVOLC and MODLEN in detecting volcanic hot spots, is also evident considering that both algorithms did not identify most documented eruptive activities of OL, because of common issues generally affecting the identification of thermal anomalies from space (about 13% of total days with documented eruptions were detected by satellite). In more detail: • Clouds: About 36% of analyzed MODIS scenes were completely overcast over the OL summit area, as indicated by the OCA method. As an example, because of cloud coverage both RSTVOLC and MODLEN undetected the explosive eruption of 4 September 2007, which was presumably Despite the differences mentioned above, the similar behavior of RST VOLC and MODLEN in detecting volcanic hot spots, is also evident considering that both algorithms did not identify most documented eruptive activities of OL, because of common issues generally affecting the identification Remote Sens. 2018, 10, 1177 13 of 17 of thermal anomalies from space (about 13% of total days with documented eruptions were detected by satellite). In more detail: • Clouds: About 36% of analyzed MODIS scenes were completely overcast over the OL summit area, as indicated by the OCA method. As an example, because of cloud coverage both RST VOLC and MODLEN undetected the explosive eruption of 4 September 2007, which was presumably accompanied by a strong thermal activity in the crater and by lava flows to the west and northwest [32]. • Short-lived eruptions (e.g., strombolian paroxysms; lava fountains): These events, common at OL [31], may be undetected from space due to timing of satellite overpasses [57]. In addition, in our case, we analyzed only nighttime scenes further reducing the probability of detecting short-lived events occurring in daylight conditions. • Satellite viewing geometry: Weak thermal activities (e.g., minor intra-crater eruptions) generating hot spots of low temperature and/or spatial extent are more difficult to identify from space when satellite data are acquired under unfavorable viewing conditions (i.e., high zenith angles) (e.g., [7]).
In spite of the aforementioned issues and even considering the intrinsic limitations of used methods (e.g., RST VOLC was less effective than MODLEN in detecting hot spots during March-April because of residual cloudy pixels contaminating the spectral reference fields), this study shows that satellite systems are very important for the surveillance of OL. Indeed, the analysis of MODIS observations have revealed that several OL thermal activities were unreported by volcanological bulletins. Among the undocumented eruptive events, the identification of a phase of thermal unrest occurring during 22-24 September 2002 is particularly relevant. Indeed, the common thermal anomaly detection in that period strength the hypothesis that a short-term preparatory phase preceded the highly fluid pahoehoe lava from the lower northern slope of the monitored volcano or that the lava effusion took place before 26 September 2002, when it was observed on the ground [41].

Conclusions
The identification of subtle hot spots is an open challenge in different fields of applied remote sensing, such as volcano activity monitoring or fire detection.
In this paper, we have analyzed the thermal activity of Oldoinyo Lengai, which is an open system highly sensitive to external forcing (e.g., atmospheric pressure, earth tides [46]). Because this volcano is so unusual and so remote, it is particularly interesting to investigate by satellite.
Performance of the RST VOLC multi-temporal algorithm in detecting natrocarbonatite lava events have been assessed by comparison with MODLEN, which may be considered as the benchmark for this remote volcano, analyzing nine years (i.e., 2000-2008) of nighttime MODIS scenes, corresponding to 4754 orbits (including both Terra and Aqua satellites).
As first, we found that RST VOLC flagged 172 volcanic thermal anomalies in about 3.6% of analyzed satellite scenes, whereas according to MODLEN 165 thermal anomalies affected the same dataset (a good correlation coefficient R 2 = 0.87 between the two methods was also found). Moreover, about 92.2% of MODLEN and 97.7% of RST VOLC detections were ascribable to real thermal volcanic activities, leading to a very low false positive rate ranging from 2.3% (RST VOLC ) to 7.8% (MODLEN).
These results show that RST VOLC , which was designed to perform in whatever geographic area, is capable of performing similarly to MODLEN, although it was not tailored to OL conditions. Indeed, it identified subtle hot spots characterized also by values of the radiant flux around 1 MW and provided information about an additional number of undocumented eruptive activities of OL effectively integrating MODLEN detection (e.g., 115 undocumented eruptive days were identified by combining their outputs).
Hence, despite factors as cloud coverage and time duration of eruptive events strongly affecting the satellite monitoring of OL, the operational usage of RST VOLC could contribute to a better surveillance of this remote volcano even considering its easy implementation on multi-platform satellite system. RST VOLC runs, in fact, also on geostationary satellite data [58], which are the most suited to detect short-lived eruptive events [59,60]. Furthermore, it is ready to be implemented on VIIRS (the Visible Infrared Imaging Radiometer Suite), providing information in MIR and TIR bands with a higher spatial resolution (i.e., 375 m) than MODIS, enabling a better subtle hot spot identification at Lengai.
Author Contributions: T.L., F.M. and M.K. wrote the majority of the paper; M.L. and T.L. generated the RST VOLC outputs; M.K. provided the MODLEN products; N.P. and V.T. contributed in writing paper and interpreting results.
Funding: This research received no external funding.