A novel approach for vegetation classification using UAV-based hyperspectral imaging

Abstract The use of unmanned aerial vehicle (UAV)-based spectral imaging offers considerable advantages in high-resolution remote-sensing applications. However, the number of sensors mountable on a UAV is limited, and selecting the optimal combination of spectral bands is complex but crucial for conventional UAV-based multispectral imaging systems. To overcome these limitations, we adopted a liquid crystal tunable filter (LCTF), which can transmit selected wavelengths without the need to exchange optical filters. For calibration and validation of the LCTF-based hyperspectral imaging system, a field campaign was conducted in the Philippines during March 28–April 3, 2016. In this campaign, UAV-based hyperspectral imaging was performed in several vegetated areas, and the spectral reflectances of 14 different ground objects were measured. Additionally, the machine learning (ML) approach using a support vector machine (SVM) model was applied to the obtained dataset, and a high-resolution classification map was then produced from the aerial hyperspectral images. The results clearly showed that a large amount of misclassification occurred in shaded areas due to the difference in spectral reflectance between sunlit and shaded areas. It was also found that the classification accuracy was drastically improved by training the SVM model with both sunlit and shaded spectral data. As a result, we achieved a classification accuracy of 94.5% in vegetated areas.

classification accuracy was drastically improved by training the SVM model with 37 both sunlit and shaded spectral data. As a result, we achieved a classification 38 accuracy of 94.5% in vegetated areas. aerial survey with multispectral cameras over a citrus orchard, and found that 57 UAV-based datasets yielded better classification accuracy than aircraft-based 58 datasets for diseased citrus trees. Moreover, Peña et al. (2013Peña et al. ( , 2015 applied 59 However, because the number of mountable imaging sensors is limited by the 65 payload weight capacity of the UAV, it is difficult to obtain in-depth spectral 66 information using this type of multispectral imaging system. Hence, we adopted 67 a liquid crystal tunable filter (LCTF) for a UAV-based hyperspectral imaging 68 system. The LCTF is an optical band pass filter whose center wavelength is  Advanced spectral imaging systems on UAVs can achieve highly efficient data acquisition. However, large amounts of complex vegetation information are 83 gathered on a daily basis, and this data can be difficult to process using 84 conventional processing techniques. Recently, machine learning (ML) has 85 become one of the most powerful approaches for examining such complex 86 datasets (so-called big data). In general, ML is a data analysis method that can 87 be used to discover underlying structures, similarities, or dissimilarities present 88 in big data. In the case of supervised learning, the ML model is trained by user 89 inputs so that it gains experience throughout the training process. The ML is 90 widely applicable for identification, detection, classification, quantification, 91 estimation, and prediction in precision agriculture (Singh et al., 2016).

92
The UAV-based hyperspectral imaging system described in this study can 93 realize aerial images with a resolution on the order of tens of millimeters. Thus, it 94 could be useful for leaf-scale plant disease detection when combined with the 95 ML approach. Therefore, this new survey platform using LCTF technology will 96 make a significant contribution to future precision agriculture research. In this 97 paper, we present a UAV-based high-resolution vegetation classification map, where obj(λ) and eva(λ) are the measured spectral radiance of the target 138 object and the EVA mat, respectively, and eva(λ) is the reflectance of the EVA mat. Note that the spectral reflectances measured by the LCTF imager in this 140 study were calculated using this equation.  In the case of aerial snapshot hyperspectral imaging by the LCTF imager, the 156 captured area shifts slightly from image to image, which is due to small attitude perturbations of the UAV. In order to obtain a spectral data cube, it is necessary 158 to precisely overlap the time-sequential hyperspectral images using an image 159 processing technique. In this study, we applied Speeded-Up Robust Features   was trained with the 420 sunlit spectral reflectance data presented in Figure 4(a).

214
During the training phase, we used 33-bands of each spectral reflectance as 215 33-dimensional information, and mapped the data onto a 33-dimensional 216 hyperplane using the SVM model. The shape of the decision boundary, which is 217 the determining factor of a classification, can be mapped by the SVM model 218 based on the selected kernel type, the kernel parameters, and the training 219 dataset. Therefore, a proper setting and training dataset should be selected to 220 classify the data at a certain high level (e.g., Huang et al., 2002). In this study, 221 the radial basis function (RBF) kernel was selected, and the cost parameter 222 and the gamma of the RBF kernel were tuned to 312 and 0.5, respectively, in order to minimize the cross-validation estimate of the test set error.

225
As mentioned above, Figure 5

239
In this study, however, shaded and non-shaded pixel data were processed by a 240 common data processing method without shadow masking. In Figure 4