Neural Network for Surface Current Trajectory
Modeling From RADARSAT-1 SAR Data
2.0 Neural Network Algorithm
Figure 1 shows the basic configuration of NN used in this study, consisting of a network of the
interconnected nodes in three layers. The training network was set as hidden layer nodes to compute the
surface current movements. The input of the to the NN are the RADARSAT-1 intensity values and final
the output being with the surface current velocities and direction. The NN has only one hidden layer,
which has been proven to be sufficient for representing any complex problem. Inputs fed to the input
layer moved to the hidden layer through weighted connection and then added to a bias. The product is
then fed to a non-linear function (termed as, activation function) to modify it. The output from the hidden
layer undergoes the same procedure giving the network result at the output layer. This process of
obtaining final network output at output node from the applied inputs could be expressed mathematically
(Iwata et al., 1989) as in equation 1.
Figure 1. Neural Network Configuration
where, s is the slope parameter, used to adjust the slope of the activation function and adjusted for a
particular desired output from the input data. The network out put of the current velocities were
compared with in-situ current measurements was feed forward to update the coefficients (e.g. weights
and biases) in the network. This process of simulating the input of value of image intensity for outputs is
called training or learning of the network and continues until the error decreases to desired value. Since
each node in the network has different coefficients, the network can describe any nonlinear relationship.
A total of 100 samples of in-situ current measurements and intensity values of RADARSAT-1 image
were chosen for modeling surface current from RADARSAT-1 SAR data. Thirty samples are selected for
training the network and the rest 70 samples were used to validate data for performance of the trained
model. Sorting the input sample set in descending order of the RADARSAT-1 SAR intensity values and
in situ current measurements did selection of training data.
The surface current velocity outputs are scaled to the range [0.2 to 0.72 m/s] to match the range of
network output. Weights and biases in the network were initialized with small random numbers upon
start of training process. The training error is monitored with mean square error. It was found that, the
NN over fits the training values if the training is performed until the mean square error flattens down.
Hence, the mean square error of the training-testing set was estimated in each pass of the training and the
training is terminated as soon as it becomes the minimum.
