AN IMPROVED FEATURE EXTRACTION METHOD BASED on CONTEXT FEATURES for MULTI-SPECTRAL REMOTE SENSING IMAGERY

Feature extraction methods of multi-spectral remote sensing images is of great significance for remote sensing image analysis, but it still faces some challenges. The ability of traditional feature extraction methods based on artificial features or shallow machine learning have some shortcomings and limitations. Recently, a series of proposed R-CNN networks, especially Faster R-CNN, have achieved excellent results in the field of target recognition. However, Faster R-CNN for multi-spectral imagery object detection has several drawbacks: (1) the object spectral information cannot be fully utilized in Faster R-CNN used to process RGB images; (2) the spatial semantic relationship information which could not be mined by Faster R-CNN among remote sensing image features can improve the feature extraction ability of the network; (3) objects occupy relatively few pixels because of the low resolution of multi-spectral images, and Faster R-CNN has poor detection performance for small objects. To address the above problems, we propose an effective and novel object detection method for multi-spectral images with small objects. First, we design a feature extractor by adopting a 3D convolution neural network which can simultaneously extract spectral information and spatial information. Secondly, an object relation module for mining context information is introduced into the network. Finally, in order to solve the problem of small targets, a multi-scale object proposal network for generating regions of objects from several intermediate layers is used. We conducted a set of controlled trials on the satellite imagery feature detection dataset released by Dstl on the Kaggle website and the results showed that our approach was very effective.