Three Volcanoes In Guatemala
A Two-Year Survey of Lava Emissions
Using an Artificial Neural Network
for
V. Fuego, V. Pacaya, and V. Santa Maria
by
Disclaimer
The information provided here is presented in the hope that others may find it interesting and useful. While great care has been taken to assure correctness, completeness, and accuracy, the provided information is presented as experimental with no guarantee of suitability for any purpose.
Credits
A very special thank you goes to NASA's Direct Readout Laboratory for their support. Without their help, this project would not have been possible.
Introduction
This document outlines the results of an initial test of a recently created artificial neural network, created for the purpose of detecting lava emissions. This is part of a broader effort aimed at using artificial intelligence concepts to detect volcanic activity in general.
The Data
The data is from the Landsat 8 Collection 1, OLI/TIRS Tier 1 data sets consisting of eleven bands, quality data, and two plain text files containing metadata. The data sets were acquired using the USGS Earth Explorer as .tar.gz archives containing the data as multiple files in GEOTIFF format. The time span is December 26, 2015, through December 31, 2017. In all, 43 daytime data sets were acquired for this project. Three scenes were omitted from this study because they do not meet Tier 1 specification. The WRS2 row and path (020 050) was the same for all of the data sets used in this study.
Most of the scenes were sufficiently cloud-free as to obtain useful results. However, a few were extremely cloudy, and no useful data could be obtained from them. All available daytime scenes for the time period of this analysis were included, and evaluated, regardless of cloud cover.
The three volcanoes being studied here are well known, and there is a significant amount of online information regarding their classification, typical style(s) of erupting, as well as historic reports on activity.
This project relies significantly on multiple online sources for information, and due credit is given to those sources.
Information on the naming, geolocation, and classification of these volcanoes as used in this document comes from the Smithsonian Global Volcanism Program database.
Links to all of the online information sources used here are provided where appropriate.
The Volcanoes
Fuego
Latitude 14.473 Longitude -90.88
“Volcán Fuego, one of Central America's most active volcanoes, is one of three large stratovolcanoes overlooking Guatemala's former capital, Antigua. The scarp of an older edifice, Meseta, lies between 3763-m-high Fuego and its twin volcano to the north, Acatenango. Construction of Meseta dates back to about 230,000 years and continued until the late Pleistocene or early Holocene. Collapse of Meseta may have produced the massive Escuintla debris-avalanche deposit, which extends about 50 km onto the Pacific coastal plain. Growth of the modern Fuego volcano followed, continuing the southward migration of volcanism that began at Acatenango. In contrast to the mostly andesitic Acatenango, eruptions at Fuego have become more mafic with time, and most historical activity has produced basaltic rocks. Frequent vigorous historical eruptions have been recorded since the onset of the Spanish era in 1524, and have produced major ashfalls, along with occasional pyroclastic flows and lava flows.” (Global Volcanism Program, 2013. Fuego (342090) in Volcanoes of the World, v. 4.6.4. Venzke, E (ed.). Smithsonian Institution. Downloaded 13 Jan 2018 (https://volcano.si.edu/volcano.cfm?vn=342090&vtab=GeneralInfo))
Pacaya
Latitude 14.382 Longitude -90.601
“Eruptions from Pacaya, one of Guatemala's most active volcanoes, are frequently visible from Guatemala City, the nation's capital. This complex basaltic volcano was constructed just outside the southern topographic rim of the 14 x 16 km Pleistocene Amatitlán caldera. A cluster of dacitic lava domes occupies the southern caldera floor. The post-caldera Pacaya massif includes the ancestral Pacaya Viejo and Cerro Grande stratovolcanoes and the currently active Mackenney stratovolcano. Collapse of Pacaya Viejo between 600 and 1500 years ago produced a debris-avalanche deposit that extends 25 km onto the Pacific coastal plain and left an arcuate somma rim inside which the modern Pacaya volcano (Mackenney cone) grew. A subsidiary crater, Cerro Chino, was constructed on the NW somma rim and was last active in the 19th century. During the past several decades, activity has consisted of frequent strombolian eruptions with intermittent lava flow extrusion that has partially filled in the caldera moat and armored the flanks of Mackenney cone, punctuated by occasional larger explosive eruptions that partially destroy the summit of the growing young stratovolcano.” (Global Volcanism Program, 2013. Pacaya (342110) in Volcanoes of the World, v. 4.6.4. Venzke, E (ed.). Smithsonian Institution. Downloaded 13 Jan 2018 (https://volcano.si.edu/volcano.cfm?vn=342110&vtab=GeneralInfo))
Santa Maria
Latitude 14.757 Longitude -91.552
“Symmetrical, forest-covered Santa María volcano is one of the most prominent of a chain of large stratovolcanoes that rises dramatically above the Pacific coastal plain of Guatemala. The 3772-m-high stratovolcano has a sharp-topped, conical profile that is cut on the SW flank by a large, 1.5-km-wide crater. The oval-shaped crater extends from just below the summit to the lower flank and was formed during a catastrophic eruption in 1902. The renowned plinian eruption of 1902 that devastated much of SW Guatemala followed a long repose period after construction of the large basaltic-andesite stratovolcano. The massive dacitic Santiaguito lava-dome complex has been growing at the base of the 1902 crater since 1922. Compound dome growth at Santiaguito has occurred episodically from four westward-younging vents, the most recent of which is Caliente. Dome growth has been accompanied by almost continuous minor explosions, with periodic lava extrusion, larger explosions, pyroclastic flows, and lahars.” (Global Volcanism Program, 2013. Santa Maria (342030) in Volcanoes of the World, v. 4.6.4. Venzke, E (ed.). Smithsonian Institution. Downloaded 13 Jan 2018 (https://volcano.si.edu/volcano.cfm?vn=342030&vtab=GeneralInfo))
Method
The data was classified using an artificial neural network written in the C programming language and trained for the purpose of identifying molten Lava.
The neural network is a feed forward design. It uses nine of the Landsat 8 bands in a pixel-by-pixel spectral classification.
The neural network was trained with a limited set of Tier 1 data using back propagation in a supervised classification. The training data was molten basaltic lava pixel data from multiple Landsat 8 scenes, across nine bands. The Landsat 8 data sets were used without additional processing prior to presenting it to the neural network for training or classification.
The neural network classifier produces a 30m/pixel resolution, 16 bit unsigned integer data mask with each pixel set to a value ranging from zero (0) to 255. A zero value indicates no classification, while classified pixels are encoded with a range from 1 to 255. The resolution and geolocation information is the same as for for the reflective bands as shown in the metadata files for the Landsat 8 data set.
The neural network classifies relative confidence that the spectral and thermal response of the pixel in question matches the values for molten lava from the training set. The color classification used in the browse images is a red-orange-yellow gradient in 255 steps. Red is considered “low confidence”, orange is “medium confidence”, and yellow, is “high confidence”.
The neural network was not trained to detect fire. However, the spectral and thermal characteristics of fire, as a class, overlaps with those for lava, allowing for the detection of fire.
Each scene was classified using the 30m/pixel resolution data, and the resulting classification data was interpolated to a 15m/pixel mask using a nearest neighbor interpolation. The classification data was then projected onto a 15m/pixel resolution true color browse image which was then converted to jpeg format to produce a browse scene as an aid to visual analysis.
Analysis
Each of the color browse scenes were evaluated individually and interpreted visually. Links to selected cropped 15m/pixel browse images of activity are provided for each scene evaluation. In the case of heavy cloud cover and no lava classification, only scenes with clearly identifiable ground features are shown (there is no point in showing pictures of clouds). Visual indications of ash are reported in the analysis when clearly present. Only images of representative fires are shown (there are just too many fires to show them all). Links to the 30m/pixel classification mask data (in compressed gzip format) are provided. A discussion of the results appears at the end of the analysis section.
Here is a link to an unclassified 15m/pixel browse image for the entire scene in JPEG format. It is 16384x16384 resolution. This is a 20MB download. A computer with at least 8 GB ram is suggested if you intend to view this scene.
Reports of activity for Fuego are HERE.
Reports of activity for Pacaya are HERE.
Reports of activity for Santa Maria are HERE.
December 26, 2015 LC08_L1TP_020050_20151226_20170224_01_T1
Fuego; Lava detection , medium to high confidence. An ash cloud is drifting westerly.
Pacaya: Low to medium lava detection.
Santa Maria; low to medium confidence lava detection. An ash cloud is visible.
Many smoky fires appear across the southern regions of the scene.
Classification data: LC08_L1TP_020050_20151226_20170224_01_T1.MASK.gz
January 11, 2016 LC08_L1TP_020050_20160111_20170224_01_T1
No lava detection at Fuego and Pacaya. Fuego shows some visible ash production.
Low confidence lava detection at Santa Maria.
Smoky ground fires are to be seen in various locations across the scene.
Classification Data: LC08_L1TP_020050_20160111_20170224_01_T1.MASK.gz
January 27, 2016 LC08_L1TP_020050_20160127_20170224_01_T1
Fuego is displaying medium and low confidence lava in the crater, and a lava flow down the southeastern side of the mountain. A thin cloud of ash can be seen coming from the crater, drifting eastward.
Pacaya is showing detected lava in the crater, and a modest output of ash.
Santa Maria is obscured by clouds.
A few smoky ground fires are visible in the southern regions of the scene.
Classification data: LC08_L1TP_020050_20160127_20170224_01_T1.MASK.gz
February 12, 2016 LC08_L1TP_020050_20160212_20170224_01_T1
Lava extends out of the crater onto the southeast flank of Fuego. Detected lava scoring medium to high confidence. No noticeable ash output.
Pacaya continues to show medium and low confidence lava detected at the crater. A thin ash emission can be seen drifting to the west.
No lava detection at Santa Maria. The crater is obscured by clouds.
A number of smoky fires are present.
Classification data: LC08_L1TP_020050_20160212_20170224_01_T1.MASK.gz
February 28, 2016 LC08_L1TP_020050_20160228_20170224_01_T1
Fuego and Pacaya are obscured by clouds (no images). No lava was detected.
At Santa Maria, there is low confidence (deep red) lava detection extending from the crater, to the east. A slight amount of ash output is visible.
A few ground fires are visible through openings in the cloud cover.
Classification data: LC08_L1TP_020050_20160228_20170224_01_T1.MASK.gz
March 15, 2016 LC08_L1TP_020050_20160315_20170224_01_T1
Medium and high confidence lava detected at Fuego. The lava flow on the southeast flank continues.
Pacaya Obscured by clouds and no lava detected.
Santa Maria is obscured by clouds (no image). No lava detection.
Classification data: LC08_L1TP_020050_20160315_20170224_01_T1.MASK.gz
March 31, 2016 LC08_L1TP_020050_20160331_20170223_01_T1
This scene is very hazy, with lots of cloud cover, making it difficult to see ash clouds and smoke from fires.
Fuego is visible, and showing medium and high value detection. The lava flow on the southeast flank of the mountain remains about the same in appearance as in previous weeks.
Pacaya is obscured by clouds. No lava detection.
Santa Maria is partly visible. There is low confidence lava detection.
Fires are present.
Classification data: LC08_L1TP_020050_20160331_20170223_01_T1.MASK.gz
April 16, 2016 LC08_L1TP_020050_20160416_20170326_01_T1
This scene is very hazy, with significant cloud cover.
At Fuego, two lava flows are visible adjacent to the crater.
Lava scoring low confidence (deep red) values.
Pacaya is obscured by clouds (no image).
Santa Maria is obscured by atmospheric haze (no image).
Only a few fires are present.
Classification data: LC08_L1TP_020050_20160416_20170326_01_T1.MASK.gz
May 18, 2016 LC08_L1TP_020050_20160518_20170324_01_T1
This is the most spectacular scene in this data set.
Fuego erupted in grand style, with a large ash cloud visible drifting to the southwest.
The volcano is partially obscured by ash emissions. There is a long lava flow (partly obscured by clouds and ash) extending from the area of the crater, down the southeast flank of the mountain. Detected lava scored from lowest to highest confidence values in this scene.
Pacaya is visible. Medium confidence lava pixels were detected. A modest emission of ash is visible in this scene.
Santa Maria is obscured by clouds (no image).
Only a few smoky fires in this scene.
Classification data: LC08_L1TP_020050_20160518_20170324_01_T1.MASK.gz
June 3,2016 LC08_L1TP_020050_20160603_20170324_01_T1
All three volcanoes are obscured by clouds. No lava pixels were detected (no images).
There is a dark brown ash cloud extending above the clouds at Fuego.
Much of the scene is relatively cloud-free. No fires were detected.
Classification data: LC08_L1TP_020050_20160603_20170324_01_T1.MASK.gz
June 19, 2016 LC08_L1TP_020050_20160619_20170323_01_T1
Fuego with low to high confidence lava detection at , and around the crater. Lava is overflowing the summit crater by a short distance. Ash emissions drifting northwest.
Pacaya is obscured by cloud cover. No lava detection.
Santa Maria showing low confidence lava detection.
No fire detection in this scene.
Classification data: LC08_L1TP_020050_20160619_20170323_01_T1.MASK.gz
July 5, 2016 LC08_L1TP_020050_20160705_20170323_01_T1
Fuego with high confidence lava detection in the crater, and medium confidence detection of lava around the crater. An ash cloud is visible, drifting southwest.
Pacaya showing low to medium confidence lava detection.
Santa Maria shows low confidence lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20160705_20170323_01_T1.MASK.gz
July 21, 2016 LC08_L1TP_020050_20160721_20170323_01_T1
Mostly cloudy/overcast.
Fuego displays low-medium confidence lava detection at the crater.
Pacaya is obscured by clouds (no image). No lava detection.
Santa Maria is obscured by clouds (no image). No lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20160721_20170323_01_T1.MASK.gz
August 6, 2016 LC08_L1TP_020050_20160806_20170322_01_T1
Mostly cloudy. All three volcanoes are obscured by clouds. No lava detection. No images.
No fires detected.
Classification data: LC08_L1TP_020050_20160806_20170322_01_T1.MASK.gz
August 22, 2016 LC08_L1TP_020050_20160822_20170322_01_T1
Fuego obscured by clouds (no image). No lava detection.
Pacaya showing low-medium confidence lava detection at the crater.
Santa Maria is obscured by clouds (no image). No lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20160822_20170322_01_T1.MASK.gz
September 7, 2016 LC08_L1TP_020050_20160907_20170321_01_T1
Fuego, with a long lava flow to the southeast. Medium to full confidence detection values. Ash emissions drifting to the west.
Pacaya is obscured by clouds. No lava detection.
Santa Maria is obscured by clouds. No lava detection.
Several fire detections in urban areas.
Classification data: LC08_L1TP_020050_20160907_20170321_01_T1.MASK.gz
September 23, 2016 T1 data not available
October 9, 2016 LC08_L1TP_020050_20161009_20170320_01_T1
Obscured by clouds. No lava detection.
No fire detection.
Classification data: LC08_L1TP_020050_20161009_20170320_01_T1.MASK.gz
October 25, 2016 LC08_L1TP_020050_20161025_20170318_01_T1
Mostly obscured by clouds. No lava detection. No images.
No fires detected.
Classification data: LC08_L1TP_020050_20161025_20170318_01_T1.MASK.gz
November 10, 2016 LC08_L1TP_020050_20161110_20170318_01_T1
Fuego, with low confidence (deep red) lava detection overflowing the crater.
Pacaya obscured by clouds. No lava detection.
Santa Maria Obscured by clouds. No lava detection.
Fires detected.
Classification data: LC08_L1TP_020050_20161110_20170318_01_T1.MASK.gz
November 26, 2016 LC08_L1TP_020050_20161126_20170317_01_T1
Fuego displaying ash and steam production. No lava detection.
Pacaya obscured by clouds.
Santa Maria obscured by clouds.
A few scattered fires.
Classification data: LC08_L1TP_020050_20161126_20170317_01_T1.MASK.gz
December 12, 2016 LC08_L1TP_020050_20161212_20170316_01_T1
Fuego, High confidence lava detection in the crater with medium confidence lava overflowing the crater rim.
Pacaya, medium to high confidence detection at the crater. There is a little bit of ash production drifting westerly.
Santa Maria obscured by clouds. No lava detection.
An increasing number of fires.
Classification data: LC08_L1TP_020050_20161212_20170316_01_T1.MASK.gz
December 28, 2016 LC08_L1TP_020050_20161228_20170314_01_T1
Fuego, with high confidence lava detection at the crater. Low confidence lava detection overflowing the crater with a lava flow to the southwest. An ash cloud drifts westerly.
Pacaya obscured by clouds.
A nice clear view of Santa Maria. No lava detection.
More, and larger smoky fires.
Classification data: LC08_L1TP_020050_20161228_20170314_01_T1.MASK.gz
January 13, 2017 LC08_L1TP_020050_20170113_20170311_01_T1
Fuego with low to high confidence lava detection. Slight amount of visible ash output.
Pacaya, with low to medium confidence lava detection. No ash output visible.
Santa Maria; No lava detection. Visible ash output.
Many smoky fires.
Classification data: LC08_L1TP_020050_20170113_20170311_01_T1.MASK.gz
January 29, 2017 LC08_L1TP_020050_20170129_20170214_01_T1
Fuego showing low confidence lava detection on the southwest side of the crater. Ash output clearly visible.
Pacaya obscured by clouds. No lava detection.
Santa Maria with low confidence lava detection in the crater.
Classification data: LC08_L1TP_020050_20170129_20170214_01_T1.MASK.gz
February 14, 2017 LC08_L1TP_020050_20170214_20170228_01_T1
Fuego; Medium to high confidence lava detection. High confidence detection in the crater, low and medium confidence detection around the crater.
Pacaya obscured by clouds.
Santa Maria obscured by clouds.
Many smoky fires.
Classification data: LC08_L1TP_020050_20170214_20170228_01_T1.MASK.gz
March 2, 2017 LC08_L1TP_020050_20170302_20170316_01_T1
Fuego; Low to medium confidence lava detection. Lava flowing from the crater down the southeastern slope.
Pacaya; Medium to high confidence lava detection overflowing the crater in all directions.
This is the biggest, brightest lava emission detected by the neural network for Pacaya for the two years that this analysis spans.
Santa Maria; Mostly obscured by clouds. No lava detection.
Classification data: LC08_L1TP_020050_20170302_20170316_01_T1.MASK.gz
March 18, 2017 LC08_L1TP_020050_20170318_20170328_01_T1
Fuego; Completely covered by clouds. Low to medium confidence lava detection.
Pacaya; Completely cloud covered, no lava detection, no image.
Santa Maria; Mostly cloud covered, no lava detection.
A few fires are detected in this scene.
Classification data: LC08_L1TP_020050_20170318_20170328_01_T1.MASK.gz
April 3, 2017 LC08_L1TP_020050_20170403_20170414_01_T1
Fuego; Low to high confidence lava detection. Lava overflowing the crater and down the southeastern slope of the mountain.
Pacaya; Low to medium confidence lava detection.
Santa Maria; Partly obscured by clouds. No lava detection. Some ash production.
Many fires detected in the southern areas of the scene.
Classification data: LC08_L1TP_020050_20170403_20170414_01_T1.MASK.gz
April 19, 2017 LC08_L1TP_020050_20170419_20170501_01_T1
Fuego; Low to high confidence lava detection. Lava overflowing the crater and spilling onto the immediate vicinity.
Pacaya; Low to medium confidence lava detection at the crater.
Santa Maria; Partly obscured by clouds. No lava detection.
Classification data: LC08_L1TP_020050_20170419_20170501_01_T1.MASK.gz
May 5, 2017 LC08_L1TP_020050_20170505_20170515_01_T1
Obscured by clouds. No lava detection. No images.
No fires detected.
Classification data: LC08_L1TP_020050_20170505_20170515_01_T1.MASK.gz
May 21, 2017 LC08_L1TP_020050_20170521_20170526_01_T1
Fuego; Low to high confidence lava detection. Lava is overflowing the west side of the crater. Ash output drifting to the west.
Pacaya; Completely obscured by clouds and haze. No lava detection. No image.
Santa Maria; Completely cloud-covered. No lava detection. No image.
Partly cloudy in the southern agricultural areas. No fires detected.
Classification data: LC08_L1TP_020050_20170521_20170526_01_T1.MASK.gz
June 06, 2017 LC08_L1TP_020050_20170606_20170616_01_T1
Cloudy, and hazy near the volcanoes.
Fuego. A cluster of five low-medium confidence 30m pixels appears northwest of the crater location (approx lat, lon: 14.481, -90.897). This may be a pyroclastic flow that was partially visible to the spacecraft through an opening in the clouds. There is considerable activity for this time period, with reports of pyroclastic flows and lava flows.
Pacaya; Crater obscured by clouds, no lava detection.
Santa Maria; Cloud-covered. No lava detection. No image.
Large, cloud-free areas in the south. No fires detected.
Classification data: LC08_L1TP_020050_20170606_20170616_01_T1.MASK.gz
June 22, 2017 LC08_L1TP_020050_20170622_20170630_01_T1
Fuego; The crater is completely cloud-covered. There is medium and high confidence detection of lava. It looks as though Fuego has blown a "hole" through the cloud cover.
Pacaya; Completely obscured by clouds. No lava detection. No image.
Santa Maria; Completely obscured by clouds. No lava detection. No image.
Many large partly cloudy and clear areas. No fires detected.
Classification data: LC08_L1TP_020050_20170622_20170630_01_T1.MASK.gz
July 8, 2017 LC08_L1TP_020050_20170708_20170716_01_T1
Fuego; Obscured by clouds. Low confidence lava detection at the crater extending to the southeast.
Pacaya: Visible, but no lava detection.
Santa Maria; Mostly obscured by clouds. No lava detection.
Clear to partly cloudy. No fire detection.
Classification data: LC08_L1TP_020050_20170708_20170716_01_T1.MASK.gz
July 24, 2017 LC08_L1TP_020050_20170724_20170809_01_T1
Fuego; Medium to high confidence lava detection at, and around, the crater.
Pacaya; A single medium confidence lava pixel was detected in the crater.
Santa Maria; Completely cloud-covered. No lava detection. No image.
Partly cloudy. No fires detected.
Classification data: LC08_L1TP_020050_20170724_20170809_01_T1.MASK.gz
August 9, 2017 LC08_L1TP_020050_20170809_20170824_01_T1
Fuego; There is a long lava flow from the crater down the southwest flank of the mountain. Lava pixels are medium to highest confidence. A dark ash cloud extends to the north from the crater.
Pacaya; Visible, no lava detection.
Santa Maria; Cloud-covered. No lava detection. No image.
No fires detected.
Classification data: LC08_L1TP_020050_20170809_20170824_01_T1.MASK.gz
August 25, 2017 LC08_L1TP_020050_20170825_20170913_01_T1
Obscured by clouds.
No images.
No lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20170825_20170913_01_T1.MASK.gz
September 10, 2017 LC08_L1TP_020050_20170910_20170927_01_T1
Obscured by clouds.
No images.
No lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20170910_20170927_01_T1.MASK.gz
September 26, 2017 T1 data not available
October 12, 2017 LC08_L1TP_020050_20171012_20171024_01_T1
Obscured by clouds. No lava detection.
No fires detected.
Classification data: LC08_L1TP_020050_20171012_20171024_01_T1.MASK.gz
October 28, 2017 T1 data not available
November 13, 2017 LC08_L1TP_020050_20171113_20171122_01_T1
Fuego erupting! A thick ash emission drifting to the south from the crater.
High and medium confidence lava detection is shown on the north side of the crater.
Pacaya; No lava detection.
Santa Maria; Obscured by clouds. No lava detected.
Fires detected. Quite a few smoky fires in the south.
Classification data: LC08_L1TP_020050_20171113_20171122_01_T1.MASK.gz
November 29, 2017 LC08_L1TP_020050_20171129_20171207_01_T1
Fuego; Medium and high confidence lava detected at, and around the crater.
Pacaya; A single-pixel low confidence lava detection in the crater.
Santa Maria; Obscured by clouds. No lava detection.
Classification data: LC08_L1TP_020050_20171129_20171207_01_T1.MASK.gz
December 15, 2017 LC08_L1TP_020050_20171215_20171224_01_T1
Fuego; Medium and high confidence lava detection at, and around, the crater.
Pacaya; Medium and high confidence lava detection at, and around the crater.
Santa Maria; No lava detection.
Classification data: LC08_L1TP_020050_20171215_20171224_01_T1.MASK.gz
December 31, 2017 LC08_L1TP_020050_20171231_20180104_01_T1
Fuego; Medium to high confidence lava detected. There is a dark ash cloud drifting north.
Pacaya; Low to medium confidence lava detected.
Santa Maria; No lava detected.
Many fires detected at low to high confidence. Some have oversaturated pixels.
Classification data: LC08_L1TP_020050_20171231_20180104_01_T1.MASK.gz
Discussion
There are 14 officially listed volcanoes in this scene. Of those, only Fuego, Pacaya, and Santa Maria were observed to be producing lava by the neural network.
All three of the volcanoes examined here are considered to be continuously active, but Fuego is clearly the most active.
The neural network frequently detects lava through ash clouds, thin clouds, and moderate haze. This appears to reduce the confidence level of the classification in such cases. Very thick ash clouds, heavy obscuring cloud cover, and extremely hazy conditions frequently results in “no detection”.
The timing of the satellite pass is important. If the pass occurs hours or days after lava is extruded, the classification may result in no classification , or classification at a lower confidence level.
Sensor oversaturation in multiple reflective bands can result in a pixel not being classified by the neural network. Because of a neural network's ability to generalize, a single band being oversaturated may cause the neural network to classify that pixel at a lower confidence level. Sensor oversaturation can thus result in an unclassified “hole” within a group of pixels that otherwise contains high confidence pixels, or result in a lower confidence "inlier" cluster of pixels surrounded by a cluster of pixels with higher confidence values.
The thermal bands in the Landsat 8 T1 data sets have scan line lengths with valid data which are shorter than that for the reflective bands. The length of the scan line with non-zero thermal data is variable from one data set to the next. Since data beyond the scan line length is encoded as "no data" (a value of 0), the neural network will only classify a pixel when the thermal data is non-zero. The end result is that there is a narrow region along the eastern edge of the scene that is not processed by the neural network and no classification for that area can be made.
The degree of classification confidence of pixels located in the crater of an active volcano is very often much higher than for classified pixels immediately outside the crater, suggesting a temperature difference. Furthermore, the dacite lava at Santa Maria consistently classifies with a low-medium confidence level, and never at a high confidence level. Conversely, the basaltic lavas at Fuego and Pacaya are far more likely to classify with medium and high confidence values. It is well known that dacitic lavas are at lower temperatures than basaltic lavas.
Considering that the neural network was trained on basaltic lavas only, it seems consistent and correct that the lavas at Fuego and Pacaya often classify at a high confidence, while the lava at Santa Maria consistently classifies at low-medium confidence levels.
With the above in mind, It seems likely that the output produced by the neural network might be correlated to surface temperature. This has not been done as of this writing, but may be visited at a later date.
Fires are typically detected at a low to medium confidence, but many do classify as high confidence. Nearly all of the detected fires appear in the southern agricultural regions in the scenes. The occurrence of fires in Guatemala clearly varies seasonally and appear to be most common from about September through May. For both years, there are no fires detected by the neural network in the months of June, July, and August. For both years there was no detection of fire or lava in the month of October, due to poor visibility.
Detection of fire sometimes occurs in urban areas without obvious indications of smoke or burn scars. Urban ground cover is complex, and likely contains many spectral classes which could trigger a detection by the neural network. Also, in a large urban area, the likelihood of a burning building, or industrial activity being detected by the neural network is arguably high.
The tendency of the neural network to generalize solutions gives it an unavoidably qualitative aspect. Since this version of the neural network is simple, it lacks the ability to interpret the context in which a pixel classification occurs, and a human is thus required to correctly interpret the resulting data. Future versions of this AI will incorporate some ability to interpret data context so that fire can be better distinguished from lava by the AI, allowing it to make more robust qualitative decisions.
Long Term Goals
The addition of a greater diversity of relevant data to improve context interpretation should be pursued further. The addition of data sets containing information on known volcanoes and the addition of seismic data are suggested first steps. A later addition of a deep convolutional neural network for visual identification of volcanic features is logistically difficult due to the scale of the problem, but remains as a worthy goal. The identification of surface temperature, local low-altitude volcanic ash emissions, as well as gaseous emissions are also desirable. Lastly, combining all of the aforementioned ideas using an additional neural network to form a powerful neural structure that is both "deep" and "wide" has great appeal.
Conclusion
Overall, the neural network performed well, and as expected. Further testing on a larger number and wider variety of scenes should be done to assure that the neural network is consistent, and stable in it's behavior as a classifier of volcanic lava.
By itself this neural network lava classifier, while useful when interpreted by a human, is not fully adequate as a fully autonomous solution for the gross identification of volcanic activity. But, by carefully combining it with multiple sources of information, it could be developed into a powerful autonomous tool.
