**4. Results and interpretations**

212 Remote Sensing – Applications

Table 3. Error matrices of outlines obtained by photo-interpretation, comparing those from the photo-interpretation of aerial photographs with those of the different types of satellite data.

In order to validate the outlines automatic extraction method, each extraction result has been compared to the referenced outline. The error matrixes (table 4) represent the tests run on the seven tested objects localized on figure 7.

The error matrixes enable calculating global errors of classifications between 77% and 96% (table 4). We can observe a disparity between the summit zone eruptions and those implemented on sloppy substrates with various spectral properties. The first ones show a global precision between 77% and 88%, whereas the others variate from 91% to 96%. This is partly due to low reflectances observed for the substrate at the summit of the volcano. For example, the outlines of a lava flow newly implemented are hardly distinguished from the intra Dolomieu lava effusion zone that presents similar ages. The mean precisions show the same disparity, with values included between 77% and 84% at the summit zone, and included between 86% and 90% for the lava flows situated on the cone's flanks and on the slopes. The area of the lava flows also play a role in the classification precision's difference,

Automatic Mapping of the Lava Flows at

automatic extraction.

precise the "lava flow" class will be.

interact on the precision.

Piton de la Fournaise Volcano, by Combining Thermal Data in Near and Visible Infrared 215

Mean Accuracy 83.8% Producer accuracy lava flow 78.2% Producer accuracy no lava flow 89.4% Mean distance between the outlines 26m **20070218** Lava Flow class No lava flow class

Global Accuracy 77.98% Mean Accuracy 77.9% Producer accuracy lava flow 80.588% Producer accuracy no lava flow 75.22% Mean distance between the outlines 55.14m

**Lava Flow Aerial picture** 0.2960 0.0713 0.3673 **No Lava Flow Aerial picture** 0.0862 0.2617 0.3479 **Total Area of aerial picture** 0.3822 0.333 0.7152

Table 4. Error matrices for the seven outlines obtained by automatic extraction, by comparing those from the photo-interpretation of aerial photographs with those from

the summit zone lava flows have less superficy than the ones observed along the slopes. An error on the outline will echo as importantly as the aria of **the mapped** lava flow is low. As to say, the more the mapped lava flow's area is important, the less the error will echo on the precision of the classification. The exactitude of the producer of the "lava flow" class is the most important in the validation of an automatic lava flow outline extraction method. The later varies from 74% and 87% on the entire tested lava flow. There is no observed difference between the eruptions that occur in the Dolomieu crater and the ones that occur on the flanks and the Grandes Pentes of the volcano. On another hand, the morphology of the lava flows plays a role on this precision (figure 8). The more the lava flow's superficy is important and the more compact it is, like a cercle shape form or very large, the more

The distances between the automatic extracted lava flow outlines, in the case of superimposed lava flows, and those considered as reference is about 30m (table 4 and figure 9). Now, the satellite data used in the automatic extraction have a visible, near and medium infrared spatial resolution from 10m to 20m. If the extraction was realized without any possible confusion, we should have mean distance of about the size of these pixels. Nevertheless, in some low reflectances zones, only the thermal imaging is able to bound the outline's localization zone (figure 8 and 9). It therefore considerably reduces the precision because this data has a 90m pixel. We notice that in zones where the lava regularly effuses, the reflectance's difference cannot allow a free from error extraction of the outline, and it is therefore deducted from the thermal mask. This mask is essential because it allows the definition of a zone of interest inside witch the outline can be determinated. There is no

In vegetalized zones, the extraction can be considered as a simple threshold of the infrared, the outline is then about the size of the pixel. In some zones, especially the Dolomieu crater or certain parts of the cone, the differences of reflectances of the lava flows don't allow to differentiate them. The outline is then obtained thanks to the thermal data. It can represent up to 40% of the summit zone lava flows outline, but only 5% to 0% of the proximal and distal eruption's outlines. The mean error observed on the outline's extraction can then reach twice the size of the optical data's pixel. Some light diffusion and wedging effects can

possible confusion: to one thermal data, only one lava flow can be associated.


**Lava Flow Aerial picture** 3.422 0.659 4.081 **No Lava Flow Aerial picture** 0.667 11.602 12.269 **Total Area of aerial picture** 4.089 12.261 16.35

**Lava Flow Aerial picture** 2.620 0.425 3.045 **No Lava Flow Aerial picture** 0.696 9.039 9.735 **Total Area of aerial picture** 3.316 9.464 12.78

**Lava Flow Aerial picture** 0.2386 0.0445 0.2831 **No Lava Flow Aerial picture** 0.0784 0.3439 0.4223 **Total Area of aerial picture** 0.317 0.4223 0.7054

**Lava Flow Aerial picture** 1.8389 0.6246 2.4635 **No Lava Flow Aerial picture** 0.6610 26.2154 26.8319 **Total Area of aerial picture** 2.4999 26.84 29.2949

**Lava Flow Aerial picture** 3.291 0.901 4.192 **No Lava Flow Aerial picture** 0.816 17.002 17.818 **Total Area of aerial picture** 4.107 17.903 22.01

**Lava Flow Aerial picture** 0.2727 0.0761 0.3488 **No Lava Flow Aerial picture** 0.1462 1.2272 1.3734 **Total Area of aerial picture** 0.4189 1.3033 1.7222

**20001012** Lava Flow class No lava flow class

Global Accuracy 91.89% Mean Accuracy 89.205% Producer accuracy lava flow 83.85% Producer accuracy no lava flow 94.56% Mean distance between the outlines 28.41m **20010611** Lava Flow class No lava flow class

Global Accuracy 91.228% Mean Accuracy 89.445% Producer accuracy lava flow 86.04% Producer accuracy no lava flow 92.85% Mean distance between the outlines 28.54m **20030530** Lava Flow class No lava flow class

Global Accuracy 82.58% Mean Accuracy 82.855% Producer accuracy lava flow 84.28% Producer accuracy no lava flow 81.43% Mean distance between the outlines 30.45m **20040502** Lava Flow class No lava flow class

Global Accuracy 95.31% Mean Accuracy 86.09% Producer accuracy lava flow 74.64% Producer accuracy no lava flow 97.70% Mean distance between the outlines 31.55m **20050217** Lava Flow class No lava flow class

Global Accuracy 92.2% Mean Accuracy 86.965% Producer accuracy lava flow 78.51% Producer accuracy no lava flow 95.42% Mean distance between the outlines 29.65m **20051004** Lava Flow class No lava flow class

Global Accuracy 87.1%


Table 4. Error matrices for the seven outlines obtained by automatic extraction, by comparing those from the photo-interpretation of aerial photographs with those from automatic extraction.

the summit zone lava flows have less superficy than the ones observed along the slopes. An error on the outline will echo as importantly as the aria of **the mapped** lava flow is low. As to say, the more the mapped lava flow's area is important, the less the error will echo on the precision of the classification. The exactitude of the producer of the "lava flow" class is the most important in the validation of an automatic lava flow outline extraction method. The later varies from 74% and 87% on the entire tested lava flow. There is no observed difference between the eruptions that occur in the Dolomieu crater and the ones that occur on the flanks and the Grandes Pentes of the volcano. On another hand, the morphology of the lava flows plays a role on this precision (figure 8). The more the lava flow's superficy is important and the more compact it is, like a cercle shape form or very large, the more precise the "lava flow" class will be.

The distances between the automatic extracted lava flow outlines, in the case of superimposed lava flows, and those considered as reference is about 30m (table 4 and figure 9). Now, the satellite data used in the automatic extraction have a visible, near and medium infrared spatial resolution from 10m to 20m. If the extraction was realized without any possible confusion, we should have mean distance of about the size of these pixels. Nevertheless, in some low reflectances zones, only the thermal imaging is able to bound the outline's localization zone (figure 8 and 9). It therefore considerably reduces the precision because this data has a 90m pixel. We notice that in zones where the lava regularly effuses, the reflectance's difference cannot allow a free from error extraction of the outline, and it is therefore deducted from the thermal mask. This mask is essential because it allows the definition of a zone of interest inside witch the outline can be determinated. There is no possible confusion: to one thermal data, only one lava flow can be associated.

In vegetalized zones, the extraction can be considered as a simple threshold of the infrared, the outline is then about the size of the pixel. In some zones, especially the Dolomieu crater or certain parts of the cone, the differences of reflectances of the lava flows don't allow to differentiate them. The outline is then obtained thanks to the thermal data. It can represent up to 40% of the summit zone lava flows outline, but only 5% to 0% of the proximal and distal eruption's outlines. The mean error observed on the outline's extraction can then reach twice the size of the optical data's pixel. Some light diffusion and wedging effects can interact on the precision.

Automatic Mapping of the Lava Flows at

article, and to know their precision.

**6. Acknowledgements** 

**7. Note** 

Reunion website.

**8. References** 

KALIDEOS program (http://kalideos.cnes.fr).

of Economy, Trade and Industry) programm. Thanks to Rebecca Roger for the translation.

Geophysical Research, n°96 (B1), pp. 475-484.

by using coherence data.

methodology.

Piton de la Fournaise Volcano, by Combining Thermal Data in Near and Visible Infrared 217

gain of time. The gaps of observations due to clouds, or zone with high thermal diffusion and low reflectances, could be filled by using RADAR imaging in treatment sequences (Weisseil et al.2004). However, by adding Bi data will increase the treatment time, especially

The association of thermal and optical data has already been realized in other automatic classifications with outlines extractions contexts: glaciology (Raciviteanu et al. 2008), the canopy (Joshi et al., 2006), agriculture (Kasdan, 1979; Saito et al., 2001). For similar spectral resolution data, the error matrix results are comparable to those obtained by our

This methodology was developed in order to automate lava flow outline's extraction and therefore ensure a fast update of the Piton de la Fournaise's database. The lava flow map was updated thanks to photo-interpretation and automatic extraction (figure 4). It allowed us to test the reliability of the outlines extracted according to each methodology used in this

The errors measured by the matrixes give us the extracted surface's error, by comparing the automatic lava flow's area and the referenced one. We saw that: 1/ the difference of interpretation between two operators for the same data can be of 2% for the aerial photos and it varies between 2% and 5% with satellite data. 2/if the outlines extracted from satellite data's photo interpretation are compared to those extracted from aerial data, the exactitude producer of the "lava flow" class can reduce to 80%, but is at an average of 90%. 3/ Those obtained with the same type of data varies from 74% to 87%, which is to say an average of less than 10%. It influences the obtained area in less than 10%, exception made for low scale summit zone eruptions which represent an error of 20%, which remains modest and without

major consequences on the volume and production estimation rates for the volcano.

These reasearch were financed by the "Region Reunion" and the ministry of "l'Enseignement Supérieur et de la Recherche" program. The authors would like to thank the CNES (Centre National d'Etudes Spatial) for the free access to the SPOT data via the

This piece of work also used free ASTER data obtained thanks to AIST GEO Grid (Ministry

All the lava flow outlines' vectors will be available online on the Laboratoire de Geosciences

Abrams M., Abbott E., KahleA., 1991. Combined Use of Visible, Reflected Infrared, and

Thermal Infrared Images for Mapping Hawaiian Lava Flows. Journal of

Fig. 9. Automatic lava flow outline extraction of May 2004 (blue) compared with the referenced outline (purple).

Lets considerate the example of the May 2004 lava flow (figure 9 and table 4). Three zones are sharply distinguished in the extraction: a very low reflectance and high thermal zone (1), a low reflectance with few or no thermal diffusion zone (2), and a various reflectance with few or no thermal diffusion zone (3). The weakest precision is for the first zone, for the second zone, the low reflectance is due to the substrate's nature and to luminosity issues in the zone, because the shadow projected by the rampart can interfere. A luminosity parameter is to be taken into consideration when choosing and time acquiring SPOT data. As for the extraction of the third zone, the thermal infrared band essentially obtains it, because the lava flow implemented in a vegetalized zone with a high spectral signature difference compared to the lava flow. It is the zone where the extraction is the most precise because only based on the SPOT data.

#### **5. Discussion and conclusion**

The vectors of the outlines obtained by automatic extraction properly match with the referenced outlines since we obtain a mean exactitude producer of 80% for the "lava flow" class. If we compare our automatic extraction results to those obtained by DGPS, they are less precise. The DGPS has a precision of about one centimeter at its antenna, now it's the operator who transports it. The error is due to the positioning of the antenna regards to the outlines. The error then is about one meter, for pluri-meter in our extractions. Nevertheless, the effusion zones are not all accessible, and cannot let realizing the outlines in their integrality, especially in the Grandes Pentes. Computer assisted drawing methods are on the other hand less reliable than our methodology for the data' distortion are not taken into account and can provoke hundred of meters error. The results of our classification are close to those obtained in other contexts' literature. (Azerzaq et al., 1997 ; Messar et Messar, 1997 ; Yüksel et al., 2008).

The vectorization time by photo-interpretation can take several days, whereas the automatic method enables obtaining this vector in less than one hour. There is therefore a considerable gain of time. The gaps of observations due to clouds, or zone with high thermal diffusion and low reflectances, could be filled by using RADAR imaging in treatment sequences (Weisseil et al.2004). However, by adding Bi data will increase the treatment time, especially by using coherence data.

The association of thermal and optical data has already been realized in other automatic classifications with outlines extractions contexts: glaciology (Raciviteanu et al. 2008), the canopy (Joshi et al., 2006), agriculture (Kasdan, 1979; Saito et al., 2001). For similar spectral resolution data, the error matrix results are comparable to those obtained by our methodology.

This methodology was developed in order to automate lava flow outline's extraction and therefore ensure a fast update of the Piton de la Fournaise's database. The lava flow map was updated thanks to photo-interpretation and automatic extraction (figure 4). It allowed us to test the reliability of the outlines extracted according to each methodology used in this article, and to know their precision.

The errors measured by the matrixes give us the extracted surface's error, by comparing the automatic lava flow's area and the referenced one. We saw that: 1/ the difference of interpretation between two operators for the same data can be of 2% for the aerial photos and it varies between 2% and 5% with satellite data. 2/if the outlines extracted from satellite data's photo interpretation are compared to those extracted from aerial data, the exactitude producer of the "lava flow" class can reduce to 80%, but is at an average of 90%. 3/ Those obtained with the same type of data varies from 74% to 87%, which is to say an average of less than 10%. It influences the obtained area in less than 10%, exception made for low scale summit zone eruptions which represent an error of 20%, which remains modest and without major consequences on the volume and production estimation rates for the volcano.
