**3. Method**

204 Remote Sensing – Applications

transformation of the rock on the surface, or because of the implantation (always very fast in Reunion Island) of a vegetation cover (lichen, moss, shrub…). Lava flows which implementations were separated by several years can therefore be distinguished by their spectral properties. For the summit zone of the volcano, all the more in the Dolomieu crater, where the rocks are superimposed with only a few years or a few months of interval, the

The oldest known eruption at the Piton de la Fournaise dates back to 1644. About 200 events have been counted since that date thanks to archives, 95% of them took place in the caldera (Lacroix, 1936; Stieltjes et al., 1989; Peltier et al. 2009). Never the less, this database is incomplete, particularly in the case of short time and low scale eruptions, before 1980. The mean average magma emission at Piton de la Fournaise estimated over a century, is 0,01 km3.an-1 (Lénat et Bachèlery, 1987), or 0.3 m3.s-1. The debit estimations show a temporal estimation. For example, Stieltjes et al. (1989), calculate a mean debit of 0.3 m3.s-1 over 54 years (1931-1985), but obtain 0.78 m3.s-1 for a period of 25 years (1960-1985). These variations are partly due to the existence of long periods of inactivity. For example, no eruption took place during 1992 and 1998; witch is to say 6 years of inactivity. Also, another inactivity as long was observed between 1966 and 1972 (Villeneuve, 2000). Peltier et al. (2009) illustrate these debits variation and show a more important activity since 1998. Between January 1990 and January 2010, 61 eruptions have been registered with a total volume of emitted lava estimated at 473 Mm3 (figure 2), and 33 eruptions between 1998 and 2010, with a total volume of emitted lava of 313 Mm 3 (Peltier et al. 2009, OVPF 2009; 2010). From these observations, we have calculated a mean debit estimated between 0.45 m³.s-1 and de 0.82 m3.s-1, from 1980 to 2010. These estimations are superior to those obtained by Stieltjes et al.

Fig. 2. Estimation of the cumulative volume of lava emitted from 1980 to 2010 by the Piton

spectral properties of the diverse lava flows can then be very similar.

(1989), on former periods.

de la Fournaise.

The originality of this research states in the use of thermal data as an analyze mask. In spite of its low spatial resolution (90m), thermal data brings essential information in our automatic mapping method. It allows determining with certitude the zone where the newly implemented lava flow is localized. The automatic extraction of the outline can be realized in this analyze mask. Also, its utilization enables treating the lava flows separately from one another because for one thermal image, only one lava flow is associated in this methodology. This is particularly adapted in the case of the constitution of a lava field flow.

### **3.1 Optical and thermal data**

The automatic extraction method has been realized by the combination of thermal and optical data. SPOT and ASTER data have been used. SPOT data have **a** wavelength from the visible to mid infrared, and a spatial resolution from 2,5 to 20 meters. ASTER data have a spectrum from visible to thermal infrared and a spatial resolution that varies from 15m (VNIR) to 90m (TIR).

The ASTER TIR thermal data have to be acquired at the end of the eruption or very little time afterwards, in order for the thermal anomaly to be clearly visible on the entire zone. The maximum post eruption delay of acquisition is variable and depends on the thickness of the lava flow and therefore on its speed of cooling. In most of the cases, it is less than a month. ASTER VNIR and SPOT data can be acquired long after the lava flow's implementation. The principal is not having a new lava flow implementing on the same zone. The recent lava flows present low reflectances between the visible and mid infrared wavelength. The basalt spectrum, in the visible and short wavelength of the infrared (0.4-2.4 µm), is dominated by the presence of iron, which, at different levels of crescent oxidation, increases the reflectance (Despinoy, 2000). In the same way, the presence of lichen that grows on the lava flow increases the reflectance. In near and medium infrared, the presence of chlorophyll in the vegetation induces a strong signal (Kahle et al. 1995), permitting to discriminate precise outlines in zones with vegetation cover, especially near the Grandes Pentes area as for the Piton de la Fournaise.

The KALIDEOS project from the CNES (Centre National d'Etudes Spatiales, and GEO Grid (AIST/METI), and the NASA, grant free satellite data in the case of a research program. The data used in this article are from SPOT data from the KALIDEOS program, and from the GEO Grid program for the ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) data. The climatological conditions must be optimal because the presence of clouds masks, deforms or reduces the thermal anomaly, which in a tropical zone and especially in the volcano's zone is frequent. Only seven ASTER TIR data (the ASTER satellite was released in orbit in December 1999), acquired at the end of the eruption don't present these types of issues, and have therefore been used for the treatments (Table 1 and figure 3). Optical data acquired after the eruption have been associated to the former images.

#### **3.2 Error calculation and outlines precision of photo-interpretation**

The precision of the outlines extracted by automatic method leans entirely on their comparison with a base of outlines realized from very high-resolution satellite data' photo interpretation, and aerial pictures from IGN. This base is considerate as a reference.

Automatic Mapping of the Lava Flows at

Table 2. An error matrix for two classes

that has been classified

obtained by automatic extraction:

(http://ccrs.nrcan.gc.ca/glossary/index\_e.php?id=3124 (modifiée)).

(http://ccrs.nrcan.gc.ca/glossary/index\_e.php?id=3124).

**3.2.1 Error matrix** 

(table 2).

Piton de la Fournaise Volcano, by Combining Thermal Data in Near and Visible Infrared 207

The error matrix enables comparing two thermal maps. These matrixes are constructed according to a methodology developed by the Remote sensing Center of Canada

In our study case, an outline is compared to another that will be considered as reference

In the matrix, numbers contained by cells correspond to areas. The total of the areas by column represent the total area of the class obtained by automatic classification. The total of the areas of each line corresponds to what has been correctly classified. The area values present in the diagonal correspond to the ones correctly classified. The other cells represent the areas which were classified wrong, either by omission, or by commission (table 2).

It is therefore possible to calculate a global precision (in percent) of the classification, which is to say the area of what has been classified correctly in regards to the total area of the zone

total area of the study ∗ 100

Pmi i The precision of each class is different. Some pixels can be attributed to classes that don't suit them, if the spectral properties are similar. The exactitude producer (table 2) bases itself on the area of the pixels correctly classified in regards to the area of the lava flow considered as reference. As for the exactitude user (table 2), it bases on the area of the pixel correctly

The exactitude producer of the "lava flow" class allows knowing the accuracy of the outline

P� � <sup>∑</sup> Well class area

It is also interesting to look at the mean precision of each class of classification:

classified in regard to the area of the lava flow obtained by extraction.

Pm �

Fig. 3. Chronology of the acquisitions of the thermal data from the eruptions at the Piton de la Fournaise.


Table 1. SPOT and ASTER data used for automatic treatments.

The outlines realized after photo interpretation, also known as reference, are not without errors. They contain errors linked to the operator's subjectivity and from the resolution of the data used to extract the outlines. One has to take into consideration these imprecisions before interpreting the results. This will be approached in the photo interpretation chapter, where outlines extracted by satellite data' photo interpretation are tested and compared on the base of those extracted from IGN's aerial pictures.

The method chosen lies on the constitution of an error matrix, because it permits to know the precision of all or part of the classification. It's expressed in percent or area.

#### **3.2.1 Error matrix**

206 Remote Sensing – Applications

Fig. 3. Chronology of the acquisitions of the thermal data from the eruptions at the Piton de

Table 1. SPOT and ASTER data used for automatic treatments.

the base of those extracted from IGN's aerial pictures.

The outlines realized after photo interpretation, also known as reference, are not without errors. They contain errors linked to the operator's subjectivity and from the resolution of the data used to extract the outlines. One has to take into consideration these imprecisions before interpreting the results. This will be approached in the photo interpretation chapter, where outlines extracted by satellite data' photo interpretation are tested and compared on

The method chosen lies on the constitution of an error matrix, because it permits to know

the precision of all or part of the classification. It's expressed in percent or area.

la Fournaise.

The error matrix enables comparing two thermal maps. These matrixes are constructed according to a methodology developed by the Remote sensing Center of Canada (http://ccrs.nrcan.gc.ca/glossary/index\_e.php?id=3124).

In our study case, an outline is compared to another that will be considered as reference (table 2).

In the matrix, numbers contained by cells correspond to areas. The total of the areas by column represent the total area of the class obtained by automatic classification. The total of the areas of each line corresponds to what has been correctly classified. The area values present in the diagonal correspond to the ones correctly classified. The other cells represent the areas which were classified wrong, either by omission, or by commission (table 2).


Table 2. An error matrix for two classes

(http://ccrs.nrcan.gc.ca/glossary/index\_e.php?id=3124 (modifiée)).

It is therefore possible to calculate a global precision (in percent) of the classification, which is to say the area of what has been classified correctly in regards to the total area of the zone that has been classified

$$\text{Pg} = \frac{\sum \text{Well class area}}{\text{total area of the study}} \ast 100$$

It is also interesting to look at the mean precision of each class of classification:

$$\text{Pm} = \frac{\text{Pmi}}{\text{i}}$$

i The precision of each class is different. Some pixels can be attributed to classes that don't suit them, if the spectral properties are similar. The exactitude producer (table 2) bases itself on the area of the pixels correctly classified in regards to the area of the lava flow considered as reference. As for the exactitude user (table 2), it bases on the area of the pixel correctly classified in regard to the area of the lava flow obtained by extraction.

The exactitude producer of the "lava flow" class allows knowing the accuracy of the outline obtained by automatic extraction:

Automatic Mapping of the Lava Flows at

1974 ; Bussière, 1967 ; Mc Dougall, 1971).

**3.2.3 Automatic process: TIR- VNIR method** 

(figure 5).

Piton de la Fournaise Volcano, by Combining Thermal Data in Near and Visible Infrared 209

then less precise. The SPOT satellite was put into function in 1986, the outline data base was completed by the mappings of the OVPF and the BRGM (Stieltjes et al., 1985 et 1989 ; Billard,

Our methodology leans on different types of satellite data, it is important to know the influence of the spatial resolution on the extracted outlines. A comparison between the outlines obtained by photo-interpretation of different data and the referenced outlines is then done by using an error matrix and a mean distance between the outlines (table 3). According to the used satellite data, the awaited error on the outline is about the same size of the pixel (table 3). The classification's precision can tell that a mapping by satellite data' photo-interpretation is 85% more reliable for satellite data with a 10m to 20 m resolution, and 95% reliable for THR SPOT data (table 3). The other error due to the referenced outlines extraction can be the consequence of the operator's subjectivity. It presents an error from 2%

The lava flows which implement in zones where spectral properties are different, such as a vegetation zone or a soil that presents high spectral reflectances, are the easiest to identify

Fig. 5. Examples of automatic extraction of the lava flow outline when the reflectance of the substrate is very different from that of the lava. A and B: lava flow from the eruption of

To automatically extract an outline, different methods can be used: classifications, threshold, or automatic detection of change (Inglada et al., 2003; Habib et al., 2007). The distance between the referenced outlines and the outlines automatically extracted is then proportional to the pixel's size, and the lava flow's classification precision is between 95% and 99%. On the other hand, lava flows which implement in low spectral reflectance zones, such as the central cone of the Piton de la Fournaise or the upperstream part of the Grandes Pentes, ask for more complex treatments. A data treatment methodology is then put together by using ENVI software (Figure 6). The visible data Principal Composant Analysis (PCA) is applied in order to maximize the data's anti-correlation. The thermal bands, near

September 2009. C and D: lava flow from the eruption of 1986 off enclosures.

to 5%. The same test was run from aerial photos, and the errors didn't exceed 2%.

$$\text{Pm (lava)} = \frac{\text{Well class area}}{\text{Real lava flow area}} \ast 100$$

The mean distance separating outlines vectors is calculated from the area which was not correctly classified, divided by the smallest of the two compared lava flow outline's perimeter:

$$\text{Error } \pounds = \frac{\sum \text{Not correctly class area}}{\text{Perimeter}} \ast 100$$

#### **3.2.2 Photo interpretation**

The « referenced » outlines have been realized on a period from 1980 to 2010, by vectorization on the base of IGN aerial data from 1997 and 2003. As for the eruptions post August 2003, the research was done on a base of SPOT THR, or SPOT 4 AND 5 satellite data (Figure 4). The vectorization of the lava flows by using aerial photos is one of the most precise (Paparoditis et al. 2006), but it is also one of the most binding to realize. It could ask for several worked days.

Fig. 4. Piton de la Fournaise lava flow cartography between 1980 and 2010 (Servadio et al. 2008 modified).

For the entire French territories, the temporal recurrence in the IGN's acquisitions is 5 years. It is a problem when dealing with superimposed lava flows. Satellite imaging is then a complementary tool because, even if the spatial resolution of SPOT data is lower (2,5m to 20m with optical SPOT data, or inferior or equal to 1 meter with aerial photos delivered by IGN for the BDORTHOs), the temporal resolution grants data between eruptions and defines outlines for the superimposed zones. Unfortunately, the resolution of the outlines is then less precise. The SPOT satellite was put into function in 1986, the outline data base was completed by the mappings of the OVPF and the BRGM (Stieltjes et al., 1985 et 1989 ; Billard, 1974 ; Bussière, 1967 ; Mc Dougall, 1971).

Our methodology leans on different types of satellite data, it is important to know the influence of the spatial resolution on the extracted outlines. A comparison between the outlines obtained by photo-interpretation of different data and the referenced outlines is then done by using an error matrix and a mean distance between the outlines (table 3). According to the used satellite data, the awaited error on the outline is about the same size of the pixel (table 3). The classification's precision can tell that a mapping by satellite data' photo-interpretation is 85% more reliable for satellite data with a 10m to 20 m resolution, and 95% reliable for THR SPOT data (table 3). The other error due to the referenced outlines extraction can be the consequence of the operator's subjectivity. It presents an error from 2% to 5%. The same test was run from aerial photos, and the errors didn't exceed 2%.
