大規模ニューラルネットワークにおける学習の高速化 : 衛星画像における雲近傍効果の評価への応用

概要

論文の詳細を見る
The paper applies a neural network to correct near-cloud effect found in satellite ocean color data, which appears as a decrease in the estimated phytoplankton pigment concentration in the vicinities of bright clouds. An analytical model, which evaluates the magnitude of the secondary scattering of cloud-scattered light, was used to generate target signals to a neural network. The size of the network becomes significantly large because the cloud could affect as far as 80 pixels from the cloud edge in the imagery obtained by Nimbus-7/ CZCS, a typical satellite ocean color sensor. To solve the above problem we basically use the three-layer feedforward network, which is trained in a supervised manner with the improved error-back-propagation (BP) algorithm which was proposed by one of the authors and makes the original BP algorithm perform about 10 times faster. Large-scale neural network training problems are inherently difficult to perform so that no supervised learning strategy is feasible, and hence we will introduce two new network structures based on the property of the target signals. One is a partially connected network. The other is a multiple division of input and hidden units stepwise in the learning process. From a computer experiment, the combination of ,the structural modification performs about 16 times faster than the unmodified network structure with the improved BP algorithm.
東海大学の論文
1998-03-30