| dc.description.abstract |
Hyperspectral satellite imagery provides a wealth of spatial and spectral information about a given scene of interest. Therefore it is widely used in several applications like pixel-wise classification, vegetation mapping, ocean color monitoring and so on. Many pixel-wise classification algorithms like support vector machine, random forest, parallelopiped classifier, and neural networks are used for this purpose. The advent of convolutional neural networks (CNN) has brought about great development in this field, owing to their unique property of automatic feature extraction. Plain CNN architectures perform only one of pooling/convolution at each stage for feature extraction. This paper describes a new CNN architecture, the Inception SN, which makes use of both pooling and convolution at each stage to effectively extract features. It also makes use of spatial and spectral information in order to carry out classification. The outcome of this is a robust algorithm which performs well even with lower training data. |
en_US |