| Accurate and reliable crop classification information is a significant data source for agricultural monitoring and food security evaluation research.The rapid development of remote sensing technology provides great technical support for agricultural information acquisition.Polarimetric synthetic aperture radar(Pol SAR)is an active microwave remote sensing system with all-day and all-weather earth observation capability,which can provide abundant information for crop classification.However,the single-temporal SAR data cannot completely reflect the external forms change as the crops grow up,which has certain limitations.In addition,compared with optical data,radar images are often difficult to interpret,while optical remote sensing can obtain the multispectral characteristics of crops that are conducive to interpretation.Therefore,this paper focuses on solving the problem of crop classification based on multi-temporal radar remote sensing and multi-source data fusion.The main research achievements are as follows:(1)To solve the problem that multi-temporal polarimetric synthetic aperture radar data are prone to "dimension disaster" after polarization feature decomposition,a feature extraction method of sparse auto-encoder with non-negative constraints is proposed in this paper.This method adds non-negativity and sparsity constraints to the traditional auto-encoder.The purpose is to further increase the sparsity of hidden layer,compress the number of features,and speed up the convergence of the network training process.Experimental results show that this method can achieve better effect of dimension reduction,and the operation is faster than the typical feature extraction method.(2)Aiming at the problem that the classification effect of similar crops is poor due to the wide variety of crops,this paper constructs a novel discrimination network with multi-scale features,which uses convolutional kernels of different sizes to extract feature information of different scales,and then fuses the information of several branches into softmax classifier to complete classification.Through this network structure,small characteristic differences between crops can be extracted,which is beneficial to improve the classification accuracy of similar crops.The experimental results show that the average classification accuracy of the discrimination network with multi-scale features for spring wheat,mixed pasture,mixed hay and chemical fallow is more than 31% higher than that of the convolutional neural network.Meanwhile,the proposed classification method combining sparse auto-encoder with non-negative constraints and discrimination network with multi-scale features achieves the optimal classification performance,with the overall classification accuracy reaching 99.33% and the Kappa coefficient reaching 99.19%.(3)It may be difficult to interpret the backscattering characteristic information collected by radar remote sensing,while the multi-spectral characteristic information collected by optical remote sensing system may be affected by weather.To solve the above problems,this paper fused the SAR data obtained by Sentinel-1 system and the multi-spectral data obtained by Sentinel-2 system,and constructs a parallel convolutional neural network for classification.The feasibility of multi-source fusion data classification is proved by experiments.The final experimental results show that using multi-source data fusion improves the overall classification accuracy by 5% and Kappa coefficient by 10%compared with using a single data source. |
