摘要： 目的 几何光学四分量是指在太阳光照条件下传感器所能观测的4个光学分量，即光照植被、光照土壤、阴影植被和阴影土壤。四分量是构成遥感几何光学模型的重要内容。在近地表遥感应用中，相机俯视拍照是提取四分量的一个途径。准确快速地从图像数据中提取四分量对植被冠层结构参数反演和植被长势监测具有重要意义。方法 植被与土壤二分量的识别是四分量提取的基础。目前大多数二分类算法在自然光照条件复杂时分类误差较大。本文基于卷积神经网络(CNN)和阈值法实现了多种二分类和四分量提取算法。阈值法中，使用SHAR-LABFVC (Shadow-Resistant Algorithm: LABFVC) 实现植被与土壤的二分类，并在此基础上应用二次阈值分割获取四分量，称为二次阈值法；基于CNN的方法中，采用U-Net架构，并使用RGB和RGBV数据进行训练得到U-Net 和U-Net-V模型。最后，我们将U-Net与阈值法结合，实现了一种混合算法。 结果 我们在18张图像（1800个子图）数据上进行了实验，结果表明，与目视解译得到的四分量真值相比较，U-Net-V和混合法精度最高，具有相近的均方根误差RMSE (0.06和0.07)和相关系数（0.95和0.94）；二次阈值法与U-Net模型精度略低于上述两种算法，RMSE分别是0.08和0.09，相关系数均为0.88。在二分类实验中，U-Net的分类正确率是91%，SHAR-LABFVC为85%。结论 通过本文对比实验证明，在二分类问题中，U-Net可以更好的应对复杂自然光照条件下的数字图像。在四分量提取实验中，混合法和U-Net-V的结果优于U-Net与二次阈值法，可以用于提取四分量。
Algorithm for classifying Geometric-optical four-components from images in natural illumination condition
GaoZebin,QuYonghua(Beijing Normal University)
Abstract: Objective The four-component of Geometric-optical model, i.e., lit vegetation, lit soil, shaded vegetation, and shaded soil, could be observed by optical sensors in nature light condition. The four-component are the important parameters of Geometric-optical model. Images taken from a downward-looking canopy digital camera is an important source to derive the four-component. It is a great significance to propose a rapid and accuracy method of extracting the four-component for the application of canopy parameters inversion, such as leaf area index(LAI) and average leaf inclination angle. However, most of the algorithms only distinguish the vegetation and soil (i.e., two-class task) pixels, and it is found that the classification error is large in the condition of complex natural light. The mainly error is produced by specular reflection pixels, which are nearly to be white in image, and shadow-canopy pixels, which are nearly to be black in image. With the rapid development of deep learning, the accuracy of image semantic segmentation which mean classification of pixels is improved significantly. So the error introduced by specular reflection pixels and shadow canopy pixels may be make better in a simple way. Method This paper implements several two-class and four-component extraction algorithms based on convolutional neural network and threshold method. In the proposed methods, the SHAR-LABFVC is a threshold method and it is used to fulfill two-class classification. When the image taken in a direct light condition, the V channel data from HSV color model has a double-peak feature in histogram. So basing on the result of SHAR-LABFVC, by application of Otsu method to the V channel data, the four-component is classified. Thus, the above two-stepwise procedure is named double-threshold algorithm. Another algorithm is U-Net, which is a neural network-based method and is used to extract the two-class and four-component. We get two models based on U-Net, One is trained using RGB image data which named U-Net and another is trained using RGB-V image data which named U-Net-V. RGB-V date is image which combined RGB and V channel data of HSV. Finally, in order to make full use of the advantages of supervised and unsupervised algorithms, a hybrid method which combined U-Net and threshold method is proposed and is used to classify four-component. It is similar to double-threshold algorithm. Firstly, we use U-Net to get vegetation and soil pixels. Then, Otsu algorithm will be used to acquired four-components. Result The validation experiment is conducted using 18 images (1800 subgraphs), and the performance is evaluated using two metrics, i.e., root mean square error (RMSE) and Pearson"s r (r). The results shows that U-Net-V and hybrid are optimal and they have closely RMSE(0.06 and 0.07) and r(0.95 and 0.94); U-Net and double-threshold method have closely RMSE(0.09 and 0.08) and same r(0.88). In the two-class experiment, the classification accuracy of U-Net is 91%, and the SHAR-LABFVC is 85%. For two-class experiment, we also use F1 score to evaluate the result of U-Net and SHAR-LABFVC. The vegetation’s F1 score of U-Net is 0.87, which is 0.07 higher than SHAR-LABFVC. The soil’s F1 score of U-Net is 0.92, which is 0.03 higher than SHAR-LABFVC. Conclusion Through the comparative experiments, it is indicated that U-Net is superior to the other method in dealing with digital images under complex natural light condition in the two-class task. Compared to SHAR-LABFVC method, U-Net can classify specular reflection pixels well and it can produce more stable and accurate classification results. The well performance of U-Net in two-class task is contributed to the convolution structure, which can utilize information from local image data and construct complex features by simple features. On the contrary, threshold methods only use one threshold to classify all pixels, the error will be high when there are some pixels disturb the distribution of histogram. In the four-component extraction task, hybrid algorithm have better result than U-Net and double-threshold method for the benefit from good performance of U-Net in two-class task, while U-Net-V can produce the best results. U-Net can not get well performance directly, but by the way of adding V channel data to raw RGB image can improve performance greatly. We summary the RGB value of pixels and found that shadow leaf maybe closely to sun lit leaf in three-dimensional space. Combined with the result of confusion matrix, we think that shadow features are difficult to be learned in our data set. So we use RGB-V data to reduce difficulty of learning shadow features and got U-Net-V model. Suggestion was made that the double-threshold method is the best candidate method to extract the four-component in the condition that training samples is unavailable. For the case there are enough training samples, it is recommended to use U-Net-V to extract the four-component.