MOISTURE USING NEURAL

clangedbivalveΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

115 εμφανίσεις

DISAGGRREGATION OF
REMOTELY SENSED SOIL
MOISTURE USING NEURAL
NETWORKS

Marius P. Schamschula,

William L. Crosson, Charles Laymon,
Ramarao Inguva, and Adrian Steward

WAC/IFMIP
June 10, 2002

2

Background


Hydrological modeling


Models need to be initialized


Point wise data is not sufficient to initialize
models


Want to use remotely sensed data


Resolution of sensors is insufficient for
hydrological models

WAC/IFMIP
June 10, 2002

3

Disaggregation


We need some method of integrating
the low resolution remotely sensed data
into the hydrological model


This requires upsampling or
disaggregating the low resolution data
to high resolution

WAC/IFMIP
June 10, 2002

4

Disaggregation: Issues


We cannot create information


We must supply additional information


Due to discontinuities, the nature of the
data is not suitable to smoothing or
spline based interpolation

WAC/IFMIP
June 10, 2002

5

Disaggregation: Approaches


Bayesian statistics


Requires priors (high
-
resolution)


Linear Regression


Artificial Neural Networks (ANNs)


Require additional high
-
resolution
information


WAC/IFMIP
June 10, 2002

6

The Data: Little Washita River
Watershed


The dataset is based on the Southern
Great Plains ‘97 (SGP ‘97) field
experiment


16 over
-
flights with the Electronically
Scanned Thinned Array Radiometer
(ESTAR) instrument


One month of meteorological/hydrological
data

WAC/IFMIP
June 10, 2002

7

Additional Data Needed for
Training


The 16 over
-
flights are insufficient for
training an ANN


Additional data was derived from the
Simulator for Hydrology and Energy
Exchange at the Land Surface (SHEELS)

hydrological model and a Radiative
Transfer Model (RTM)


We will use synthetically generated
datasets

WAC/IFMIP
June 10, 2002

8

Neural Networks: Inputs


High resolution


Precipitation (1, 2
-
3, 4
-
6, 7
-
12, 13
-
24, 25
-
48, 49
-
96, 97
-
192 hours before current
time)


Soil type: sand and clay content


Vegetation water content


Upstream contributing area


Low resolution


Microwave emissivity

WAC/IFMIP
June 10, 2002

9

Neural Networks: Output


High Resolution


Top layer soil moisture

WAC/IFMIP
June 10, 2002

10

Linear Neural Network

WAC/IFMIP
June 10, 2002

11

Linear Neural Network:
“Design”


A linear ANN is not trained. The weights
can be directly determined by matrix
inversion


Design is fast: for about 1000 points,
350 time steps the process takes less
than 10 seconds

WAC/IFMIP
June 10, 2002

12

Linear Neural Network: A
“Little” Problem


The valid range of soil moisture is
between 0 and 1


The linear ANN in very wet conditions
will overshoot the maximum value


We can manually clip the maximum at 1…


Need to do better

WAC/IFMIP
June 10, 2002

13

Clipping Linear Neural Network

WAC/IFMIP
June 10, 2002

14

Clipping Linear Neural
Network: Training


Requires more computational
resources:


400 MB RAM


About a minute and a half CPU time

(G4 533 MHz)

WAC/IFMIP
June 10, 2002

15

Differences between Linear
and Clipping Linear NNs


Given that the ANN architectures differ
only in their transfer functions, we find


Weights are similar, except:


Precipitation from 2 to 3 hours before current
time


Clipping Linear ANN requires no post
processing for overshoot

WAC/IFMIP
June 10, 2002

16

Clipping Linear Neural
Network: Dry Conditions

1

0

“Truth”

1.6 km

12.8 km

WAC/IFMIP
June 10, 2002

17

Clipping Linear Neural
Network: Wet Conditions

1

0

“Truth”

1.6 km

12.8 km

WAC/IFMIP
June 10, 2002

18

Clipping Linear Neural
Network: High Resolution
Movie

WAC/IFMIP
June 10, 2002

19

Clipping Linear Neural
Network: Low Resolution Movie

WAC/IFMIP
June 10, 2002

20

Root Mean Square Error

WAC/IFMIP
June 10, 2002

21

Future Work


Refine the transfer function: asymptotic
exponential


An additional layer with transfer
functions appropriate for the various
inputs


Neighborhood interconnectivity

WAC/IFMIP
June 10, 2002

22

Conclusion


We can use an ANN to perform data
disaggregation of low resolution
remotely sensed microwave signals by
fusing in other data such as
precipitation, soil type, etc.


More information @

http://www.caos.aamu.edu/HSCaRS/