Documente Academic
Documente Profesional
Documente Cultură
819
(a)
Fig. 2. (a) EVI and (b) ∆LST time series for a pixel cleared 3. RESULTS AND DISCUSSION
in May 2007
In Fig.4(a) the annual evaluation of the results is shown. An-
nually, from 1138 ground truth cleared pixels, to be found
2.4. Validation Dataset between 27194 unflagged pixels, there is a 51.6 percent pre-
cision (847 hits over 1667 found), 74.4 percent recall and an
Redaf dataset was used for validation purposes. First, it F-Score of 0.60. However, a visual analysis of the false posi-
was used to check if the algorithm correctly detects the areas tives revealed that there are some geographical characteristics
cleared during 2013 in the study area (annual evaluation). Af- of the landscape that are not completely flagged (for exam-
terwards, using Google Earth Engine platform, we selected ple, cities, roads and extensive farming - puesteros), which
the polygons of 2013 from the Redaf dataset for the study the algorithm detects as deforestation. When cities and roads
area. Using Landsat monthly composed images we added to are included in the flagged area, and extensive farming is con-
the database the monthly information about the deforestation sidered, precision ascends to 62 percent, F-Score to 0.66 and
event. If an image for a specific month was not available recall descends to 0.71.
due to cloud cover, the number of unavailable images was The result of the monthly evaluation is shown in Fig.4(b).
registered. In this way, it was possible to check if the month First, it differentiates between pixels whose month of defor-
of deforestation reported by the algorithm agrees with the estation was inspected with or without cloud coverage. After-
month that would be reported using high resolution images wards, it shows the differences between the month detected
(monthly evaluation) and, if it is not the case, calculate the and the month reported. As can be seen, most of the pixels
difference between them. are detected with 0, 1 or 2 months of delay. In the case of
Two statistical measures were used to evaluate the perfor- the pixels whose month of deforestation was reported with
mance of the algorithm: precision - how many of the pixels two missing data due to cloud coverage, most of them are
820
temporal and a spatial pattern in the breakpoints. This model
could be used alone or to complement another model. The
main advantage of this model is that the annual deforestation
events were detected with higher accuracies with low com-
putational cost. On the other hand, its main limitation is that
the parameters of this model should be calibrated for each
study area, as it can be shown that is possible to increase the
percentage of hits modifying the parameters of the algorithm.
Fig.4(b) shows that it is important to consider the number of
lost images - due to cloud coverage - in the validation dataset.
This also shows the advantage of using MODIS median reso-
lution images instead of Landsat high resolution images.
Finally, the results obtained in this work are relevant for
the DCF area, since they provide a fast and alternative model
(a) to the traditional visual analysis carried out by the forest ser-
vice. To sum up, the proposed model provides a fast and ac-
curate automatic detection of deforestation at the DCF area.
5. REFERENCES
821