融合YOLOv5s与Criminisi算法的农业遥感图像去云方法研究

Fusion of YOLOv5s network and Criminisi algorithm to remove clouds from agricultural remote sensing images

【目的】构建融合YOLOv5s与改进Criminisi算法的农业遥感图像去云方法,为云层干扰环境下地表信息获取、地表物的解译等研究提供支持。【方法】首先使用基于容差的暗通道先验(dark channel prior,DCP)算法去除雾和部分薄云,以提升图像整体对比度与云层边缘清晰度;然后融合YOLOv5s深度学习网络进行云层区域阈值分割,实现云层蒙版的快速精确自动提取;最后通过样本块大小自适应调整策略对Criminisi算法进行改进,实现遥感图像的有效去云修复处理。通过对含不同大小云层的遥感图像进行去云试验,并利用信息熵、峰值信噪比(peak signal-to-noise ratio,PSNR)、均方误差(mean-square error,MSE)和结构相似性(structural similarity index measure,SSIM)4个指标对去云结果进行评价,以验证本研究算法的有效性。【结果】采用融合YOLOv5s和自适应样本块的改进Criminisi算法对8幅含云图像进行了修复,修复后图像的平均PSNR为21.01,平均SSIM为0.77;并对57幅模拟加云图像进行修复,其平均PSNR为28.59,平均SSIM为0.93,表明将改进Criminisi算法应用于遥感图像去云研究是可行的。在此基础上,对本研究算法的适用性以及阴影对去云效果影响的研究表明,不同大小和位置的云层干扰造成未知区域不确定度较大,对修复效果影响较为严重;阴影区域与云区域相接时存在阴影块填充,修复效果尚有待提升。【结论】融合YOLOv5s与改进Criminisi算法的去云方法可有效修复云层遮挡区域,同时保留较为真实的地表信息,可用于农业遥感信息精细感知研究。

【Objective】A cloud removal method for agricultural remote sensing images based on fusion of YOLOv5s network and improved Criminisi algorithm was constructed to provide support for acquisition of surface information and interpretation of surface objects under the interference of clouds.【Method】Fog and partial thin clouds were removed by dark channel prior(DCP)algorithm based on tolerance mechanism to improve global image contrast and sharpness of clouds edges.Then,the YOLOv5s deep learning network was fused in the threshold segmentation of cloudy area to achieve fast and accurate automatic extraction of cloud mask.Finally,the Criminisi algorithm was improved through the adaptive adjustment strategy of sample block size to achieve effective cloud removal and repair of remote sensing images.To verify effectiveness of the proposed algorithm,cloud removal experiments on remote sensing images with different sizes of clouds in different locations were carried out,and information entropy,PSNR,MSE,and SSIM were used as indicators to evaluate the algorithm.【Result】Eight haze images were repaired with the fusion of YOLOv5s network and the inproved Criminisi algorithm of adaptive sample blocks.The average PSNR of repaired images was 21.01 and the average SSIM was 0.77.Besides,57 simulated cloud images were repaired by this method with averaged PSNR of 28.59 and averaged SSIM of 0.93.It was feasible to apply the improved Criminisi algorithm for cloud removal of remote sensing images.Based on this,the applicability of the improved Criminisi algorithm and the impact of shadows on cloud removal were further explored.The cloud interference of different sizes and positions caused large uncertainty in unknown areas with significant impacts on repaired effect.There was shadow block filling when the shadow area was adjacent to cloudy areas,and the results needed improvement.【Conclusion】The areas covered by clouds can be effectively repaired with the fusion of YOLOv5s network and improved Criminisi algorithm and more realistic surface information was kept.Thus,the method can be used for fine perception of agricultural remote sensing information.