A Neural Network Sea Ice Edge Classifier for the NASA Scatterometer Sami M. Alhumaidi, W. Linwood Jones, Jun-Dong Park, Shannon Ferguson, and Michael H. Thursby Florida Institute of Technology 150 W. University Blvd. Melbourne, FL 32901 (407) 768-8000 x6219 (407) 984-8461 fax
[email protected],
[email protected] Simon H. Yueh Jet Propulsion Laboratory Pasadena, CA 91109 (818) 354-3012 Abstract. – The NASA Scatterometer (NSCAT) to be launched in August 1996 is designed to measure wind vectors over ice-free oceans. To prevent contamination of the wind measurements, by the presence of sea ice, an algorithm based on only NSCAT data is described. Results are presented for a neural network trained using dual linear polarized Ku-band backscatter measured by the SeaSat-A Satellite Scatterometer (SASS). These results demonstrate the utility of neural network classifiers to provide this ice flag. Results are presented for both multilayer perceptron (MLP) and a learning vector quantization (LVQ) neural networks. Classification skill is evaluated by comparisons with surface truth and with an independent ice-flagging algorithm.
INTRODUCTION In August 1996, the NASA Scatterometer (NSCAT) will be launched on Japan's Advanced Earth Observation Satellite (ADEOS) to measure the surface winds over the world's ice free oceans. On each side of the subsatellite track, the NSCAT will measure the ocean normalized radar backscatter (sigma-0) at three azimuths. Further, for one azimuth, the antenna is dual polarized with linear, vertical and horizontal polarization's (V-pol and H-pol). Since land, or ice within the antenna instantaneous field of view (IFOV), will contaminate the ocean wind measurement, it is necessary to identify and remove measurements so affected. For land, the process is relatively simple using land maps and knowledge of the antenna IFOV's. However, because of its dynamic nature, it is difficult to identify the extent of sea ice. Usually, the sea ice boundary is determined from a variety of satellite remote sensing data e.g., visible, infrared and passive microwave; however the non-simultaneity of these measurements with This work was sponsored under contract to the Jet Propulsion Laboratory - NASA Scatterometer Project .
ADEOS can lead to significant errors as well as increased complexity in the NSCAT data processing. This paper describes a convenient and timely technique for determining this boundary from the NSCAT data alone. The technique involves the use of a neural network algorithm for inferring sea ice cover using sigma-0's from the dual polarized antenna.
SEA ICE SIGMA-0's To train the neural network, backscatter data collected by the SeaSat-A Satellite Scatterometer (SASS) [1] were used. During the SASS experiment, dual polarized data were obtained for both the Arctic and Antarctic regions; however, because of a premature failure of the satellite, winter sea ice data were only available for Antarctica. Sigma-0's for July 13, 1978 were classified in a two step process. First, they were associated with ocean or sea ice using an independent copolarized backscatter algorithm by Yueh [2]. Next, they were compared with independent Antarctic ice classification charts from National Snow and Ice Data Center to establish credibility. An example of the sigma-0 data for one revolution is shown in Fig. 1. Data from four even revolutions were used to train the neural network, and data from seven odd revolutions were used as an evaluation set. Because of significant overlap of the SASS swaths, this method provided an excellent means for comparison. For the training set, a conservative approach was adopted whereby data from both the ocean and ice were deleted within about 100 km on both sides of the sea ice boundary. In this manner low ice concentrations (% area coverage) associated with the marginal ice zone were excluded. For the evaluation set, sigma-0 data were subjectively classified using an even more conservative criteria that identified mixed ocean/ice regions as ice. In this way, those sigma-0's classified as ocean were certain to be "not-contaminated".
NEURAL NETWORK CLASSIFIERS Neural nets have been successfully used for classifying radar backscatter [2-4] and, in particular, to classify SAR images of sea ice [3]. The two most common neural classifiers applied to sea ice classification are the Learning Vector Quantization (LVQ) and the Multilayer Perceptron (MLP) classifiers. The former employs a simple learning rule developed by Kohonon [5]; whereas the latter uses a number of learning rules of which the most common is the backpropagation rule, based on the gradient descent method of optimization. In this work, we applied the LVQ and MLP neural classifiers to classify sigma-0 measured by the SASS over Antarctica. The sigma-0 values were co-registered V-pol/Hpol pairs from the SASS Global 50-Km binned sigma-0’s provided by the Jet Propulsion Lab Physical Oceanography Distributed Active Archive Center. LVQ Classifier: First, we used a 6¥3 Self Organizing Map [5] neural net to obtain weights to initialize a LVQ net with three input neurons and six output neurons. Each of the output neurons was labeled by a known class (i.e., water or ice). The next step was to train this net with the training set developed using classification criteria as discussed above. The training data was iterated during the learning process for 10,000 epochs. Finally, the resulting net was tested with the evaluation data set which was subjectively classified. The result of the classification is shown in Table 1. MLP Classifier: A MLP neural net, with two hidden layers, was used to classify the same training data. The MLP net had three neurons in the input layer, 12 neurons in the hidden layers (six each) and two neurons in the output layer. The net was initialized with random weights and biases. Fig. 2 depicts the architecture of the MLP neural net that was trained using the backpropagation method with momentum term [6] until the total squared error reached a satisfactory level. The resulting weights and biases of the net were used to classify the evaluation data set. The result of the MLP classification is also listed in Table 1.
RESULTS AND DISCUSSION Although the MLP neural classifier was more difficult to design and stabilize, and took longer time to converge, it performed better than the LVQ classifier. In addition, the
MLP classifier was able to generalize better to data that was not presented during training (e.g., mixed ocean/ice). The MLP had two output categories, namely; ice, and ocean. Whenever the output of the net was -1 and 1, the class declared was ice and vice versa for water. However, when the net can not classify a point, the output will be 0 for both output neurons. In this case, the decision is “mixed”. We compared the neural network (NN) output with the subjective classification (SubC) for the odd revolutions (evaluation data set). Because the objective of this research is to properly classify ice-free ocean, the success criteria was defined as: a match between the NN and SubC for ocean and ice; or data classified by the NN as mixed when the SubC is classified as ice. Further, there were a small percentage (< 1%) of SASS ocean sigma-0's that were anomalous (H-pol > V-pol). For these cases, neural network outputs of mixed were also counted as successful. The rationale for this being that anomalous sigma-0's should be "flagged". Plots of the data classified as ice or mixed are presented in Fig. 4-6. In conclusion, the classification skill for the two neural networks are excellent which demonstrates the suitability of such algorithms for NSCAT sea ice-edge flagging. Once the designing and learning process is completed, the algorithm can produce reliable ice flags in real time. Future work will include additional input neurons to include multiple azimuth viewing.
REFERENCES [1] W.L. Grantham, E.M. Bracalente, W.L. Jones, and J.W. Johnson, "The SeaSat-A Satellite Scatterometer", IEEE J. Oceanic Eng., Vol OE-2, No. 2, pp 200-206, April 1977. [2] S.H. Yueh, R. Kowk, S.H. Lou and W.Y. Tsai, “Sea Ice Identification Using Dual-Polarized Ku-Band Scatterometer Data, IGARSS’96 digest (this issue), May 1996 [3] Y. Hara, R. G. Atkins, R. T. Shin, J. A. Kong, S. H. Yueh, R. Kwok, “Application of Neural Networks for Sea Ice Classification in Polarimetric SAR Images,” IEEE Trans. Geosci. Remote Sensing, Vol. 33, pp. 740-748, May 1995. [4] J. Orlando, R. Mann, S. Haykin, “Radar Classification of Sea-Ice Using Traditional and Neural Classifiers,” [5] T. Kohonen, Self-Organization and Associative Memory, 2nd ed., Springer, Berlin, 1988. [6] S. Haykin , Neural Networks A Comprehensive Foundation, Macmillan College Publishing Co., 1994.
Table 1: Classifiers Performance Classifier Type Percent Correct LVQ 96.73 MLP 98.4 Co-Pol 97.16
Hidden Layer 1 Input Layer
Hidden Layer 2 Output Layer
s0h s0v s0h q
Fig. 2: A 2-Hidden-Layer MLP Classifier
Figure 1 Example of SASS Sigma-0’s and Subjective Sea Ice Classification
