The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1348
Figure 4. Schematic diagram shows the partition of the H-
polarized brightness temperature (TbH) for the Roscommon
area. The corresponding V-polarized brightness temperatures
and the ground sampling volumetric soil moisture data are also
obtained together with the TbH data.
No.
Backpropagation Algorithm
RMSE
(%v/v)
1.
Batch Gradient with Momentum
4.86
2.
Gradient Descent with Adaptive Learning
Rate
5.34
3.
Gradient descent with momentum and
adaptive learning rate
4.88
4.
Resilient backpropagation
4.93
5.
Conjugate gradient backpropagation with
Fletcher-Reeves updates
*fmish : epochs = 2
4.82
6.
Conjugate gradient backpropagation with
Polak-Ribiere updates *fmish : epochs = 2
4.83
7.
Conjugate gradient backpropagation with
Powell-Beale restarts *finish : epochs = 2
4.83
8.
Scaled conjugate gradient backpropagation
5.77
9.
Quasi-Newton Algorithm : BFGS
3.93
10.
Quasi-Newton Algorithm :One step Secant
Algorithm
5.51
11.
Levenberg-Marquardt
4.04
Table 2. Root Mean Square Error (RMSE) of soil moisture
retrieval of various backpropagation training algorithms
7. DISCUSSION
Results obtained for RMSE are between the ranges of 3.93%v/v
to 5.77%v/v. The globally accepted accuracy requirement for
soil moisture retrieval is less than or equal to 4%v/v RMSE.
With the same backpropagation NN architecture (same number
of hidden layer, number of neurons in the hidden layer and
same bias and weights value), it is shown that the Broyden,
Fletcher, Goldfarb, and Shanno (BFGS) Quasi-Newton
algorithm is the only method able to retrieve the soil moisture
with the required accuracy although the Levenberg-Marquardt
method was close at 4.04%v/v. The scaled-conjugate algorithm
has the worst result at 5.77%v/v. Interestingly, all of the
methods produced results of a similar order of magnitude. The
standard deviation of the ranges of accuracy obtained is 0.55,
which is considered to be quite small, but for soil moisture
retrieval this figure is quite significant. This means that the
training algorithm when using backpropagation NN for soil
moisture needs to be checked carefully before deciding on the
algorithm to be used. The number of epochs required for each
of the different backpropagation ANN to converge varies. The
Conjugate Gradient backpropagation training algorithm
converged in 2 epochs with an average RMSE of 4.83%v/v.
8. CONCLUSIONS
This paper has described initial results on the best
backpropagation ANN to be used to model the relationship
between airborne microwave measurements of brightness
temperature and soil moisture content. A number of different
backpropagation ANN methods have been explored with the
main criterion for analysis being the RMSE between predicted
soil moisture content and ground truth. Only one method
produced results that meet the required 4%v/v criterion
although a second method was close. Examining the standard
deviation of the results across different training sets for the
same ANN method reveals significant variations. The next step
in the research is to explore the best ANN methods for
additional areas within the NAFE test site and to investigate
other available data that can be used to enhance the
performance of the ANNs for the prediction of soil moisture
content.
ACKNOWLEDGEMENT
The National Airborne Field Experiments have been
made possible through recent infrastructure (LE0453434
and LE0560930) and research (DP0557543) funding
from the Australia Research Council, and the
collaboration of a large number of scientists from
throughout Australia, United States and Europe.
REFERENCES
[1] J. A. Santanello Jr, C. D. Peters-Lidard, M. E. Garcia, D.
M. Mocko, M. A. Tischler, M. S. Moran, and D. P. Thoma,
"Using remotely-sensed estimates of soil moisture to infer soil
texture and hydraulic properties across a semi-arid watershed,"
Remote Sensing of Environment, vol. 110, pp. 79-97, 2007.
[2] T. Schmugge, "Applications of passive microwave
observations of surface soil moisture," Journal of Hydrology,
vol. 212-213, pp. 188-197, 1998.
[3] M. S. Dawson, M. S. Dawson, A. K. Fung, and M. T.
Manry, "A robust statistical-based estimator for soil moisture
retrieval from radar measurements," Geoscience and Remote
Sensing, IEEE Transactions on, vol. 35, pp. 57-67, 1997.