基于多源遥感特征融合与卷积神经网络(CNN)的丘陵地区水稻识别

Rice identification in hilly areas based on multi-source remote sensing feature fusion and convolutional neural networks(CNN)

为探究卷积神经网络(CNN)算法和多源遥感优选特征数据融合对丘陵地区水稻种植区的识别效果和适用性,以江西省上高县为研究区,利用Sentinel-2与GF-1遥感影像数据,对研究区晚稻种植区域进行识别。选取影像波段特征、植被指数、纹理特征及地形特征等分类特征,用分离阈值法(SEaTH)筛选出对各类别分离度较大的特征变量。基于Sentinel-2优选特征数据、GF-1优选特征数据、Sentinel-2与GF-1优选特征融合数据,使用CNN分类算法进行晚稻识别,同时用支持向量机(SVM)、最大似然法(MLC)分类算法进行对比。结果表明,Sentinel-2与GF-1优选特征融合数据在CNN分类算法下对水稻的识别效果最好,总体精度、Kappa系数分别为96.19%、0.93,结合野外调查数据进行验证,实际验证精度达94.69%。由研究结果可知,Sentinel-2与GF-1优选特征融合数据在CNN分类算法下对丘陵地区水稻识别具有较好的效果和适用性,是丘陵地区水稻遥感识别的有效手段。

To investigate the effect and applicability of the fusion of convolutional neural networks(CNN)algorithm and multi-source remote sensing preferred feature data on the recognition of rice growing areas in hilly areas,we took Shanggao County,Jiangxi province as the study area,and used Sentinel-2 and GF-1 remote sensing image data to identify the late rice growing areas in the study area.Classification features such as image band features,vegetation indices,texture features and terrain features were selected,and the feature variables with greater separation for each category were screened out as the preferred feature set using the seperability and thresholds(SEaTH)algorithm.The fusion of Sentinel-2 and GF-1 preferred features and Sentinel-2 and GF-1 preferred features were compared with the CNN classification algorithm for late rice recognition,and the support vector machine(SVM)and maximum likelihood(MLC)classification algorithms were used to compare the results.The results showed that the fusion data of Sentinel-2 and GF-1 preferred features had the best recognition effect on rice under CNN classification algorithm.The overall accuracy and Kappa coefficient were 96.19%and 0.93,respectively,and the accuracy was 94.69%when combined with the field survey data for validation.According to the research results,the fusion of Sentinel-2 and GF-1 preferred features had good effect and applicability for rice recognition in hilly areas under CNN classification algorithm,and was an effective tool for remote sensing recognition of rice in hilly areas.