Development of a Fused Ice Classification Scheme
Project Award Date: 09-20-2001
We propose to investigate automated learning techniques to generate classification heuristics that will allow ARKTOS to classify sea-ice in all areas of the Arctic, and, to develop techniques for the assimilation of the results of sea-ice classification algorithms, generating ice concentration maps for the whole Arctic region.
We propose to use the ARKTOS and ASIS pieces of software to generate training sets for a back propagation neural network. ARKTOS will segment SAR images and generate the numerical characteristics of the features in the images, together with numerical characteristics of ancillary data sets (SSMJI and climatology). ASIS will segment images that will then be classified by experts at the NIC. We will integrate the ARKTOS- and ASIS-generated results to create a set of vectors describing the characteristics of features together with their classification (old ice, first-year ice, new ice, and open water). These vectors will be divided into a training and test set, and we will train a back-propagation neural network to classify sea-ice. We will evaluate the neural network using the test set, the rule-based ARKTOS component, and the NIC ice charts.
We also propose to assimilate the results of sea-ice classification algorithms to generate concentration maps for the Arctic. We will generate profiles for a set of sea-ice classification algorithms and models, that describe the reliability of the output of the algorithm as a function of product, time, place, and temporal distance. Based on these reliability profiles, we will generate active assimilation maps. By clicking on the map, users will be able to access more than just concentration information. For example, a user will be able to see the high-level profile of the algorithm and data source that resulted in this classification (e.g. CALIVAL, SSMII), the detailed profile of the algorithm (e.g. reliability of the classification, publication/study on which this reliability is based, etc.), the actual raw data with ancillary information (e.g. resolution, date the image was taken, etc.), any assumptions made by the assimilation algorithm and what they were based on (e.g. why was it assumed that it was now melt onset?), any alternate classification results that can replace the most reliable ones and their details, reliability maps, where the concentration map is "stretched" in a third dimension indicating the reliability of a classification, and other information which will be identified as the users test beta versions of the software. We will also develop a declarative language for defining algorithm reliability profiles, which will allow easy editing of existing profiles and addition of new ones.
Primary Sponsor(s): Dept of Navy