基于多特征提取与注意力机制深度学习的高分辨率影像松材线虫病树识别

High-resolution image identification of trees with pinewood nematode disease based on multi-feature extraction and deep learning of attention mechanism

为了提高松材线虫病树的监测效率,减少其对林业生产造成的损失,提出一种基于多特征提取与注意力机制深度学习的高分辨率影像松材线虫病树识别方法。该方法首先在高分辨率遥感影像上提取松材线虫病树的光谱特征、空间特征等多特征,然后进行Relief特征选择算法,取特征权重前8个特征进行病树识别,发现选择差值植被指数DVI(difference vegetation index)、OHTA颜色模型的I2和I3分量作为病树与非病树的光谱特征较为合适,再运用DBscan空间聚类算法对光谱特征识别结果进行聚类,得到疑似病树像元集,此多特征识别方法识别病树的平均检测准确率为78.23%。以VGG(visual geometry group network)神经网络模型作为参考,建立VGG-S(simplification,即针对松材线虫病树进行简化)和VGG-A(attention module,即结合注意力机制)神经网络,并将人工判读生成的病树样本集和非病树样本集作为其训练样本。用以上两种不同的方法对疑似病树像元集进行识别,其中VGG-S平均检测准确率为82.61%,VGG-A的平均检测准确率为85.45%。结果表明,采用多特征和VGG-A相结合的方法在高分辨率遥感影像上识别松材线虫病树识别准确率更高。

In order to improve the monitoring efficiency of pinewood nematode disease trees and reduce their losses in the forestry production,a method of identifying pinewood nematode disease trees using high-resolution images based on multi-feature extraction and attention mechanism deep learning was proposed.Firstly,multi-features such as spectral features and spatial features of pine nematode diseased trees on high-resolution remote sensing images were extracted,then the Relief feature selection algorithm was performed by taking the top eigth features of feature weights for identifying diseased trees,and the Difference Vegetation Index(DVI),I2 and I3 components of OHTA color model was selected.The spectral features of diseased trees and non-diseased trees were found to be more suitable,and then the DBscan spatial clustering algorithm was applied to cluster the spectral feature recognition results to obtain the set of suspected diseased tree image elements,and the average detection accuracy of this multi-feature recognition method for identifying diseased trees was 78.23%.The VGG-S(simplification)and VGG-A(attention module)neural networks were established using the VGG(visual geometry group network)neural network model as a reference,and the diseased tree sample set and the non-diseased tree sample set generated by manual interpretation were used as the reference and non-diseased tree sample sets generated by manual interpretation were used as their training samples.The above two different methods were used to identify the suspected diseased tree image set,in which the average detection accuracy of VGG-S was 82.61%and the average detection accuracy of VGG-A was 85.45%.The results showed that the combination of multi-features and VGG-A was used to identify pine nematode disease trees on high-resolution remote sensing images with high recognition accuracy.In the experimental process,the light intensity of the image was insufficient,resulting in some forest trees cannot be identified on the image,thus leading to a high rate of missed detection.However,due to the flexibility of the high-resolution remote sensing image acquisition method,even if the spectral characteristics obtained from the experimental data are different under different shooting conditions and thus may lead to changes in the results,it is possible to quickly acquire remote sensing image data by means of aerial photography using unmanned aerial vehicles,etc.,and the images of higher quality can be achieved from the acquired image data for identification.