Comparison between the Effects of Visible Light and Multispectral Sensor Based on Low-Altitude Remote Sensing Platform in the Evaluation of Rice Sheath Blight
高效无损地评估农作物病害等级, 对于实际农业生产和研究都具有重要意义。 研究探讨了基于低空无人机遥感平台进行水稻纹枯病病害等级评估的可行性, 分析可见光与多光谱传感器的光谱响应差异及其对感病水稻光谱反射率获取的影响, 并定量对比两种传感器的病害监测效果。 实验研究区由67个不同品种的水稻小区组成, 每块小区均分为相接的纹枯病接种区和侵染区。 以大疆精灵Phantom 3 Advanced小型消费级无人机作为搭载平台, 分别搭载该无人机系统自带的可见光传感器和MicasenseRedEdgeTM多光谱传感器获取遥感影像。 同时, 通过植保专家现场调查的方式识别病害等级, 并利用Trimble公司的手持式NDVI测量仪获取实测NDVI值。 基于影像拼接、 波段叠合、 辐射校正后的预处理结果, 对可见光图像的接种区和侵染区共134个小区计算七种可见光植被指数, 即NDI（normalized difference index）, ExG（excess green）, ExR（excess red）, ExG-ExR, B*, G*, R*, 多光谱图像除上述可见光指数外再计算NDVI（normalized difference vegetation index）, RVI（ratio vegetation index）和NDWI（normalized difference water Index）三种多光谱植被指数。 将计算得到的图像植被指数与地面实测NDVI进行相关性分析, 以选取两种传感器的最优图像植被指数建立水稻纹枯病病害等级反演模型。 相关性分析结果表明, 基于多光谱传感器计算的图像NDVI与实测NDVI拟合度最高, 接种区R2为0.914, RMSE为0.024, 侵染区R2为0.863, RMSE为0.024。 对于可见光传感器, NDI与实测NDVI的相关性最好, 接种区R2为0.875, RMSE为0.011, 侵染区R2为0.703, RMSE为0.014。 比较两种传感器两种区域的同一图像植被指数与实测NDVI的一致性, 除B*外, NDI, ExR, ExG-ExR, G*, ExG, R*与实测NDVI基本属于高度相关, 在病害严重的接种区, 两种传感器对水稻纹枯病的监测效果相近, 但在病害相对较轻的侵染区, 多光谱传感器的监测更为精确灵敏。 基于多光谱图像NDVI建立的病害等级反演模型, R2达到0.624, RMSE为0.801, 预测精度达到90.04%, 模型效果良好。 而基于可见光图像NDI建立的反演模型, R2为0.580, RMSE为0.847, 预测精度为89.45%, 效果稍差。 对比分析可见光与多光谱传感器的光谱响应曲线, 可见光传感器可获取可见光范围的红、 绿、 蓝三个波段, 波段范围互相重叠, 多光谱传感器包含五个成像单元, 可独立获取从可见光到近红外的五个窄波光谱波段, 提供更加准确的光谱信息。 比较传感器获取的接种区和侵染区水稻平均反射率曲线得出, 多光谱传感器不仅在可见光波段反映了较可见光传感器更强的差异, 在红边和近红外波段差异则更加明显, 这说明专业窄波段传感器在病害监测方面较宽波段消费级传感器更有优势。 综上所述, 基于可见光与多光谱传感器的低空无人机遥感平台进行水稻纹枯病病害等级评估是可行的, 多光谱传感器精确灵敏, 可用于纹枯病的早期监测, 可见光传感器效果稍差但经济易于推广。 研究结果为病虫害防治提供决策支持, 有助于推动实现精准农业, 保障粮食安全。
Efficient and non-destructive assessment of crop disease grade is of great significance to the practical agricultural production and research. In this study, the feasibility of low-altitude UAV (Unmanned Aerial Vehicle) remote sensing platform for the disease grade assessment of rice Sheath Blight (ShB) was discussed. Then the spectral response differences of visible light sensor and multispectral sensor and their effects on the spectral reflectance acquisition of rice with ShB were analyzed. And rice ShB monitoring effects of two kinds of sensors were compared quantitively. The study area consisted of 67 rice plots with different varieties, each of which was divided into inoculation zone and infection zone. The drone was Phantom 3 Advanced, a small consumer-grade UAV made by DJI-Innovations company, and the payloads were the self-contained visible light sensor and Micasense RedEdgeTM multispectral sensor to acquire remote sensing images respectively. At the same time, the rice ShB disease grades were investigated by manual expert recognition and measured NDVI was obtained with Trimble's GreenSeeker Handheld Crop Sensor. Remote sensing images were preprocessed by image mosaic, layer stacking and radiometric calibration. A total of 134 plots in inoculation and infection zones of visible light image were used to calculate seven kinds of visible light vegetation indices, namely NDI (Normalized Difference Index), ExG (Excess Green), ExR (Excess Red), ExG-ExR, B*, G* and R*. Besides the above seven kinds of visible light vegetation indices, multispectral image was calculated by three kinds of multispectral vegetation indices additionally, namely NDVI (Normalized Difference Vegetation Index), RVI (Ratio Vegetation Index) and NDWI (Normalized Difference Water Index). The correlation between the image-based vegetation index and ground-based NDVI was analyzed, and the optimal image-based vegetation indices of the visible light and multispectral sensor were selected to establish the disease grade inversion model of rice ShB. The results of correlation analysis showed that the fitting degree of image-based NDVI and ground-based NDVI based on multispectral sensor was the highest, and R2 was 0.914 and RMSE was 0.024 in the inoculation zone, while R2 and RMSE were 0.863 and 0.024 respectively in the infection zone. As for the visible light sensor, the correlation between image-based NDI and measured NDVI was best, and R2 was 0.875 and RMSE was 0.011 in the inoculation zone, while R2 was 0.703 and RMSE was 0.014 in the infection zone. The consistencies of the same image-based vegetation index and ground-based NDVI of two kinds of sensors and two kinds of zones were compared, which revealed that NDI, ExR, ExG-ExR, G*, ExG, R* except B* were mainly highly correlated with the measured NDVI. In the inoculation zones with severe disease, the two kinds of sensors had similar effects on the detection of rice ShB, but the monitoring effect of multispectral sensor was more precise and sensitive in infection zones with relatively lighter disease. The disease grade inversion model of rice ShB established by NDVI based on multispectral sensor was effective, whose R2 reached 0.624, and RMSE was 0.801 and prediction accuracy was 90.04%. The disease grade inversion model established by NDI based on visible light sensor was slightly worse, whose R2 was 0.580, and RMSE was 0.847 and prediction accuracy was 89.45%. The spectral response curves of visible light and multispectral sensor were compared and analyzed. The visible light sensor can obtain three bands of red, green, blue in the range of visible light, and wavelength range overlaps with each other, while the multispectral sensor including five imaging units can independently obtain five narrow-band spectral bands from visible light to near infrared providing subtler spectral information. Through comparing the average reflectance curves of rice in inoculation zone and infection zone, the multispectral sensor not only reflected bigger difference than visible light sensor in the visible light band, but also represented more obvious difference in the red and near infrared band, which demonstrated that the professional narrow-band sensor had an advantage over broad-band consumer-grade sensor in the rice ShB monitoring. In conclusion, it is feasible to evaluate the disease grade of rice ShB based on the low-altitude UAV remote sensing platform with visible light and multispectral sensor. The multispectral sensor is precise and sensitive which can be used for early detection of rice ShB, and the visible light sensor is less accurate but economical and easy to popularize. The results of this study are expected to provide decision support for diseases control and be beneficial to promoting precision agriculture and ensureing food security.
基金项目：国家自然科学基金项目（31501222, 41201364, 41771463）, 中央高校基本科研业务费专项（2018JC012, 2017JC038, 2015BQ026, 2014JC008）, 国家大学生创新训练项目（201610504017）资助
张 建：华中农业大学资源与环境学院, 湖北 武汉 430070农业部长江中下游耕地保育重点实验室, 湖北 武汉 430070
张东彦：安徽大学安徽省农业生态大数据工程实验室, 安徽 合肥 230601
周新根：Texas A&M AgriLife Research and Extension Center, Beaumont, TX 77713, USA
刘小辉：安徽大学安徽省农业生态大数据工程实验室, 安徽 合肥 230601
谢 静：华中农业大学理学院, 湖北 武汉 430070
备注：赵晓阳, 1996年生, 华中农业大学资源与环境学院本科生
【1】Sumithra Muthayya, Jonathan D Sugimoto, Scott Montgomery, et al. Annals of the New York Academy of Sciences, 2014, 1324（1）: 7.
【2】Zhang Dongyan, Lan Yubin, Zhou Xingen, et al. In Proceedings of 2015 ASABE Annual International Meeting, New Orleans, Louisiana, USA, July 26-29, 2015.
【3】Mosleh M K, Hassan Q K, Chowdhury E H. Sensors, 2015, 15（1）: 769.
【4】ZHANG Jing-cheng, YUAN Lin, WANG Ji-hua, et al（张竞成, 袁 琳, 王纪华, 等）. Transactions of the Chinese Society of Agricultural Engineer（农业工程学报）, 2012, 28（20）: 1.
【5】ZHANG Dong-yan, SONG Xiao-yu, MA Zhi-hong, et al（张东彦, 宋晓宇, 马智宏, 等）. Scientia Agricultura Sinica（中国农业科学）, 2010, 43（11）: 2239.
【6】Anne-Katrin Mahlein. Plant Disease, 2016, 100（2）: 241.
【7】Matheus Kuska, Mirwaes Wahabzada, Marlene Leucker, et al. Plant Methods, 2015, 11（1）: 28.
【8】Elke Bauriegel, Werner B. Herppich. Agriculture, 2014, 4（1）: 32.
【9】Qin Zhihao, Zhang Minghua. International Journal of Applied Earth Observation & Geoinformation, 2005, 7（2）: 115.
【10】Faranak Ghobadifar, Aimrun Wayayok, Shattri Mansor, et al. Geomatics, Natural Hazards and Risk, 2016, 7（1）: 237.
【11】Zhou X, Zheng H B, Xu X Q, et al. ISPRS Journal of Photogrammetry and Remote Sensing, 2017, 130: 246.
【12】Jia Yinjiang, Su Zhongbin, Shen Weizheng, et al. International Journal of Smart Home, 2016, 10（5）: 159.
【13】Ballesteros R, Ortega J F, Hernández D, et al. Precision Agriculture, 2014, 15（6）: 579.
【14】Zhang Jian, Yang Chenghai, Song Huaibo, et al. Remote Sensing, 2016, 8（3）: 257.
【15】Sankaran Sindhuja, Maja Joe Mari, Buchanon Sherrie, et al. Sensors, 2013, 13（2）: 2117.
ZHAO Xiao-yang,ZHANG Jian,ZHANG Dong-yan,ZHOU Xin-gen,LIU Xiao-hui,XIE Jing. Comparison between the Effects of Visible Light and Multispectral Sensor Based on Low-Altitude Remote Sensing Platform in the Evaluation of Rice Sheath Blight[J]. Spectroscopy and Spectral Analysis, 2019, 39(4): 1192-1198
赵晓阳,张 建,张东彦,周新根,刘小辉,谢 静. 低空遥感平台下可见光与多光谱传感器在水稻纹枯病病害评估中的效果对比研究[J]. 光谱学与光谱分析, 2019, 39(4): 1192-1198