Classification and extraction of geographical elements based on remote sensing images(RSIs)is one of the key technologies for digital battlefield construction and intelligent battlefield perception.Actual application platforms face the problems of limited computing resources and insufficient training under small-samples,limiting the classification performance of geographical elements in deep neural networks.With tensor-train(TT)decomposition and weight sharing,the spatial-spectral convolutional long short-term memory(ConvLSTM3D)unit is extended to two tensorized structures(TTConvLSTM3D)and a lightweight Tensorized Spatial-spectral TTConvLSTM3D Neural Network(TST3DNN)is proposed for the classification of remote sensing images.The results on two hyperspectral HSI datasets show that TST3DNN only requires 0.34 MB of storage space,achieving better classification performance than that of the similar methods.
关键词
遥感图像/网络轻量化/卷积长短时记忆网络/张量分解/精细分类
Key words
remote sensing images/lightweight network/convolutional long short-term memory/tensor decomposition/fine classification
维普
万方数据