Abstract:

In this study, an improved convolutional neural network (CNN)approach is proposed to detect small sample targets from remotesensing images. The approach is designed to address the two issuesof small target samples and the unbalanced distribution of groundobject samples with respect to target detection of small samples byremote sensors. In the proposed method, first, $K$-nearest neighbor(kNN) regression is adopted to extract the features of each pointand convolution layer to construct the local neighborhood. Second,all local features are aggregated by maximum pooling layer in CNN torepresent global features. Subsequently, the full connection layerand scaled exponential linear unit (SELU) activation function areapplied to calculate the probability corresponding to each categoryfor classification. Finally, the proposed approach is tested andevaluated on hyperspectral imager remote sensing images datasets.Experimental results show that the proposed improvements to the CNNmodel fuse fully local features and result in the effectiverecognition and detection of small sample targets from remotesensing image with high accuracy while maintaining thenonlocal diffusion capabilities of information.

Key words: convolutional neural network (CNN), small sample, deep learning