Future Deforestation Scenarios

foulchilianΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 11 μήνες)

73 εμφανίσεις





Future Deforestation Scenarios


Methodology
Overview



















March 2012





Content



Methodology overview

................................
................................
................................
...................

3

Input data

................................
................................
................................
................................
....

3

Implementation in Brazil

................................
................................
................................
.................

4

Model training

................................
................................
................................
.............................

4

Results

................................
................................
................................
................................
.........

6

Conclusion

................................
................................
................................
................................
.......

9





Methodology overview


The algorithm
implemented for this study can be divided into three steps. During the first step,
a dataset of inputs and outputs is extracted to train a multilayer perceptron neural network.
The neural network is trained to infer the probability that a given pixel will
be deforested given
topological information (such as the distance to the nearest road, or the elevation) and the
state (deforested or not) of the pixels present in a given radius around the analyzed pixel.
During the second step, the well
-
trained neural ne
twork is applied to every pixel of the studied
area so as to generate a map of potential deforestation. In the third step, a given number of
pixels are selected to be deforested. The higher the potential deforestation value calculated by
the neural network
, the higher the likelihood that a pixel will be selected as deforested. The
default deforestation rate is constant, but it can also be set as a function of time. Finally, steps
two and three can be repeated several times in order to simulate the evolution

of deforestation
in time and in space.

Input data

For the initial

implementation of this
tool, only topographic data has

been included in the
model.
The following list

presents the input data that was

inclu
ded into the model:

1.

The distance to the nearest
road

2.

The distance to the nearest river

3.

The distance to the nearest city (> 1000 people)

4.

The elevation

5.

The base map (what was already deforested before the analysis with terra
-
i)

6.

The detection from the terra
-
I model.



Figure
1
, In
puts extracted from T
er
ra
-
i
:
pixels' status

within a given radius around the analyzed pixel
.


Implementation

in Brazil


As proof of concept, this methodology was applied for the area of the
Rio Branco


Porto Velho
road
segment
in Brazil
from the 1st of
January 2004 to the 10th of June 2011 and compared
with actual detections from Terra
-
i. The constant deforestation rate was set to 10,000 hectares
per 16 day period (equal to the average rate recorded by Terra
-
i in this area), and 10,000 pixels
were sample
d to train the neural network.
Figure 5

shows the potential deforestation for the
1st of January 2004, where the pixels directly adjacent to already deforested areas are the most
likely to be deforested. Likewise, the more remote a pixel is the lower its p
robability of being
deforested.


Model
training


The tool has been tested in

one of the study
are
a
s located

in Brazil

(Figure 2). F
igure 3
shows

the base map that was used as input for the
training.


Figure
2
, study area


Figure
3
:
The base map (
already existing deforestation

before analysis with
T
erra
-
i)

Once the neural network
was

trained, we tested it
s

capacity to recognize deforested pixels and
unchanged pixels

by classifying a validation dataset independent from the trainning dataset. As
shown by the following graph, the network seems
well able to

recognize both categories.


Figure
4
:
Neural network recognition of

deforested pixel
s

and

unchanged pixel
s

Finally, the neural network
was

applied to the whole area for 50 periods
with a constant rate of
15
,
000 pixels deforested during each period.


Results


Figure 5 shows

the potential deforestation

at time t=0. T
he pixels directly adjacent
to already
deforested areas are the most likely to be deforested. Likewise, the more remote a pixel is the lower its
probability of being deforested.


Figure
5
: Potential deforestation at t=0

Figure 6

shows

how deforestation is mo
deled by the algorithm.
Most

of the pixels selected as
cleared by the tool are located
arou
nd already deforested areas and around roads.


Figure
6
: Top: potential deforestation at t=30. Bottom:

potential deforestation at t=5
0


Figure
7
: Top: The base map with the predicted deforestation.

Bottom, the base map with the T
erra
-
i detection

The general patterns of deforestation are quite similar when comparing the

result
ing modeled
deforestation (Figure 7, top
)

with the actually detected deforestation (
Figure 7,
bottom
).


Some noise in the base map (individual pixels wrongly flagged as deforested) generated patches
at the top of the studied area. Additionally, some large events located at the bottom of the area

were missed by the model. These errors show that the base map is an important input and
must be created carefully, a task which can be difficult depending on the attributes of the
analyzed area.


Conclusion


The implementation of the methodology to
generate future scenarios of deforestation gave
encouraging results that compare favorably with Terra
-
i’s actually detected deforestation. The
general patterns resulting from the simulation are convincing and quite similar to the real
events. However, vari
ous improvements could be instigated. For example, the tool currently
only takes into account topographic variables. Ideally, other inputs such as administrative
information (protected areas, country), social information (small farmers, industrial
exploita
tion, community managed forest) and vegetation types (rather than a single measure
for all deforestation) should be included in the analysis as well.

Also
, the model is using a
constant rate which could be replaced by a variable rate inferre
d from the one
observed by the
T
erra
-
i models. Finally,
the sampling of pixel
s

to be defor
ested could be greatly improved. As
there is currently

no correlation between the pixel
s selected at a given date, it is
implied that
the selected
pixel
s

are randomly scattered
ove
r

the
entire
map
, a scenario

which is not realistic
even if the final cumulative result
appears

realistic.