Crop Classification Using Differential-Scattering-Feature Driven CNN for Dual-Pol SAR Images

Abstract

Crop type classification is one of the most important applications in Polarimetric Synthetic Aperture Radar (PolSAR) imagery. However, an increasing number of SAR systems with dual-polarization modes have been launched due to the cost and system constraints. Due to the limitation of the dual-polarization mode, there are seriously discounted characteristics for dual-pol SAR data sets, making it difficult to obtain satisfactory classification accuracy for dual-pol SAR images. Therefore, it is necessary to extract scattering characteristics which is more adapted to dual-pol SAR data sets. This paper introduces a new parameter based on the basic theory of H/α decomposition to measure the differential scattering characteristics of agricultural crops and proposes differential scattering feature driven conv-olutional neural network (CNN) for dual-pol SAR images. The experimental results show that the CNN classification methods proposed in this paper have achieved the highest classification accuracy. Compared with different feature combination inputs, the new parameter proposed in this paper can steadily improve the classification performance of the classifier, and the combination of H, α, θ, intensity features also achieves the best classification performance.