Passive Microwave Remote Sensing for Sea Ice Thickness Retrieval Using Neural Network and Genetic Algorithm

Over the years, global warming has gained much attention from the global com- munity. The fact that the sea ice plays an important role and has signiflcant efiects towards the global climate has prompted scientists to conduct various researches on the sea ice in the Polar Regions. One of the important parameters being studied is the sea ice thickness as it is a direct key indication towards the climate change. However, to conduct studies on the sea ice scientists are often facing with tough challenges due to the unfavorable harsh weather conditions and the remoteness of the Polar Regions. Thus, microwave remote sensing ofiers an attractive mean for the observation and monitoring of the changes of sea ice in the Polar Regions for the scientists. In this paper, we will be presenting 2 approaches using passive microwave remote sensing to retrieve sea ice thickness. The flrst approach involves the training and testing of the neural network (NN) by using data sets generated from the Radiative Transfer Theory with Dense Medium Phase and Amplitude Correction Theory (RT-DMPACT) forward scattering model. Once training is com- pleted, the inversion for sea ice thickness could be done speedily. The second approach utilizes a genetic algorithm (GA) which would perform a search routine to identify possible solutions in sea ice thickness that would match the corresponding brightness temperatures proflle of the sea ice. The results obtained from both approaches are presented and tested by using Special Scanning Microwave Imager (SSM/I) data with the aid of the sea ice measurements in the Arctic sea. In order to understand the interactions between the wave and sea ice medium, a forward scattering model based on Radiative Transfer Theory was constructed. This forward scattering model was further improved by incorporating Dense Medium Phase and Amplitude Correction Theory (RT- DMPACT) to take into account of the efiect of the closely placed scatterers in the sea ice medium. This forward scattering model formed the basis of our inverse model for the sea ice thickness retrieval process. For the NN approach, multiple pairs of data set consist of difierent sea ice parameters and thicknesses with the corresponding brightness temperatures are flrst generated using the forward scattering model. This data set will be provided to the NN to create a range of sea ice thickness proflle to be used for NN training. The training process is completed when the error generated by NN is acceptably small. After that, inversion is done by providing the brightness temperature proflles of the sea ice to obtain the corresponding sea ice thicknesses. As for GA, a pool of chromosomes representing sea ice thicknesses is created to be fed into the forward scattering model. The chromosomes are then evolved and carried forward to the next generation according to the natural selection concept, whereby the flttest candidate is more likely to survive and to reproduce. The generation and creation continues until the one of the chromosomes has been found to be suitable to be the thickness solution for a given brightness temperature proflle. 2. DATA TRAINING AND SEA ICE THICKNESS INVERSION BY NN The RT-DMPACT Model mentioned above is used to calculate the passive microwave returns in terms of brightness temperatures of vertically (TBv) and horizontally (TBh) polarized wave. The Neural Network (NN) constructed consists of an input layer, two hidden layers and an output layer. Each layer employs several neurons, which are connected to other neurons in the adjacent layer with difierent weights. The signals propagate from input layer, through hidden layers and to the output. The network is trained by the input-output data generated from the RT-DMPACT Model. The training process is carried out by changing the values of the interconnecting weights of the neurons in the layers by using Levenberg-Marquardt Algorithm (Martin H. & Mohammad B. M. 1994), according to the error generated. The weights in the NN are then changed in each iteration to reduce the error to an acceptable margin.