水体强反光环境中无人机多光谱影像的辐射一致性校正

为了提高水体强反光干扰的遥感影像信息提取准确度，该研究以柑橘树冠营养元素水平检测而采集的无人机多光谱影像为对象，对水体强反光造成的相同地物在不同影像上辐射信息不一致现象进行校正，从而提高营养元素水平检测的精度。首先对影像进行暗角校正，然后利用直方图对比度拉伸辅助SIFT（Scale invariant feature transform，SIFT）算法匹配出同名点，根据同名点的DN（Digital number，DN）值，利用RANSAC（Random sample consensus，RANSAC）构建校正模型对影像做相对辐射校正，并进行双边滤波去除噪声，最后经过辐射定标将影像DN值转化为反射率完成辐射一致性校正。为验证校正精度，选择蓝、绿、红、红边和近红外波段反射率以及GNDVI（Green normalized difference vegetation Index，GNDVI）的平均绝对误差（Mean absolute error，MAE）作为评价指标。试验结果表明，和直方图匹配相比，采用本文方法蓝、绿、红、红边和近红外波段校正后反射率的MAE分别为0.2%、0.5%、0.6%、1.7%和1.2%，GNDVI的MAE为0.3%，有效解决了水体反光造成的光谱失真问题，提高了受水体反光影响的遥感图像利用率，可为后续柑橘树冠营养元素估测提供精确的遥感数据保障。

Radiometric consistency correction of UAV multispectral images in strong reflective water environment

Abstract: Unmanned aerial vehicles (UAVs) have been one of the most important remote sensing monitoring tools for precision agriculture in recent years, due to their small size, light weight, and flexible flight control. However, the overall brightness of the UAV multispectral image is relatively low except for the reflection parts, resulting in uneven radiation between the images. The reason can be the strong reflection of some ground objects, such as water and stainless steel. It is likely to pose a serious impact on the subsequent application of UAV multi-spectral images in the precise monitoring of crop growth, quality, pests, and diseases. In this study, a radiation consistency correction was performed on the low-altitude airborne high-resolution UAVs multi-spectral remote sensing images that were collected from the detection of nutrient element levels in the citrus canopy, with an emphasis on the images of the same flight that caused by strong reflections of water bodies. Several types of relative correction (also referred to as radiometric normalization) have been widely used for the radiometric consistency correction between remote sensing images at present, such as Histogram matching, Pseudo invariant Features, and Statistic regression. Specifically, the fast and simple Histogram matching only considered the brightness difference between the images, resulting in the loss of image gray levels and distortion of the overall gray distribution. As such, the radiation characteristics of the ground objects were also missing in the original image. In Pseudo invariant Features and Statistic regression, the image statistics and manual selection of samples were adopted to establish sample sets, further to automatically select a sample set for a radiation correction model. These radiation consistency corrections were suitable for the same flight caused by the intensity of illumination. However, a sufficient high-quality sample set cannot be manually selected, due to the low visualization, average brightness, and overall image contrast, particularly for the low-illuminance image caused by the reflection of ground objects. Therefore, a radiant consistency correction was proposed for these UAVs'' multispectral images. First, the specific parameters of radiation correction were calculated, according to the calibration cloth. Secondly, a radial vignetting and compensation model was used to suppress the Dark-corner phenomenon in images. Then, the histogram contrast stretching was performed on the low-illuminance images to obtain enough feature points during the feature detection. Subsequently, the scale-invariant feature transform (SIFT) was employed to extract and match the feature points of the reference image and the image to be corrected. After that, a correction model was constructed using the random sample consensus (RANSAC) to correct the image, according to the digital number value (DN) of homologous points, where bilateral filtering was performed to remove the noise. Finally, the DN value of the image was converted into the reflectance represented by the parameters of radiation correction. The evaluation indicators were selected as the mean absolute errors (MAEs) of the Blue, Green, Red, RedEdge, NIR channels, and the Green Normalized Difference Vegetation Index (GNDVI) of 15 citrus trees in the overlapping area of adjacent images, in order to verify the accuracy of the correction. The test results show that the new radiation consistency correction presented a higher accuracy than the histogram matching. The MAEs of the Blue, Green, Red, RedEdge and NIR channels after correction were -0.2%, -0.5%, -0.6%, -1.7%, and -1.2%, respectively. The GNDVI MAEs was 0.003. Consequently, the spectral distortion can be relieved to effectively improve the utilization of remote sensing images caused by water reflections. The finding can also provide an accurate dataset for the subsequent quantitative research. Some challenges have still remained: 1)The highlights are needed to be removed in the reflection areas, such as water bodies. 2) There are also some reflective objects except water, such as stainless steel and glass. Anyway, the new radiation consistency correction can be extended to the more strong reflection in the application of UAV multi-spectral images in precision agriculture.