机载多光谱影像语义分割模型在农田防护林提取中的应用

农田防护林是农田生态系统的屏障,其健康状况的监测与评估在我国北方农田林网管理中尤为重要。本文以新疆生产建设兵团第三师51团为研究区,使用复合翼无人机CW-20搭载Micro MCA12 Snap多光谱相机获取农田防护林的多光谱影像,经辐射校正、裁剪等预处理,通过优选有效特征和模型比较,提出农田防护林提取的有效方法。首先,基于原始12波段,依据相关性系数矩阵和最佳指数因子（Optimum Index Factor,OIF）选取最优3波段和植被指数特征进行组合,构建8种农田防护林提取方案;然后,通过建立语义分割Deeplabv3+模型进行精度评价,得到最优3波段组合6（波长710 nm）、8（波长800 nm）、 11（波长900 nm）波段为最佳特征组合;最后,以最优3波段为基础,将Deeplabv3+模型与U-Net、ENVINet5模型进行对比分析。结果表明：Deeplabv3+模型能够更深层次的挖掘光谱中潜在的信息,相比其他模型,能够较好地处理正负样本不均衡问题,获得最高MIoU值85.54%,比U-Net、ENVINet5的MIoU值则分别高出21.21%、27.19%。该研究结果可为基于多光谱遥感影像的语义分割在农田防护林提取及健康状况监测的应用提供借鉴和参考。

Application of Airborne Multispectral Image Semantic Segmentation Model in Farmland Shelterbelt Extraction

Farmland shelterbelt is the barrier of a farmland ecosystem. Monitoring and assessing the health condition of farmland shelterbelt are significant for farmland forests management in North China. This paper selects the farmland shelterbelt in the 3^th Division of Xinjiang Production and Construction Crops as the experimental example. In July and August of 2019, we used the CW-20 fixed-wing drone which was equipped with a SONY-A7RII camera and a Micro MCA12 Snap camera to capture visible light images and multispectral images. Four targets on the ground with different reflectance (3%, 22%, 48%, and 64%) were set up for radiometric correction. we collected field data by Aowei software. Optimal bands were selected base on the Optimum Index Factor (OIF) and correlation coefficient. By integrating appropriate vegetation indices, eight different extraction schemes were constructed to select an effective method for final farmland shelterbelt extraction. In this study, when the flight mission was completed, we collated the exported POS and image data correspondingly. High resolution visible light images with 0.08 m resolution and high resolution multispectral images with 0.149 m resolution were first mosaiced using Pix4Dmapper. Four ground targets were then used for radiometric correction of mosaiced images using linear fitting model. The final experimental data was extracted using clipping method. Because each multispectral image was displayed in gray scale, which was not easy for direct data interpretation. High-resolution visible light images and multispectral images were registered together through ArcGIS software. We labeled 8030 samples based on visible light images through ENVI 5.5 software. We separated the training dataset and verification dataset according to the ratio of 4:1. Specifically, 112 512 pixel×512 pixel slices were taken as the training dataset, and 28 512 pixel×512 pixel slices were taken as the verification dataset. Based on eight extraction schemes, model accuracy of Deeplabv3+semantic segmentation model was evaluated. Three optimal bands that had the best model performance were selected including band 6 (wavelength 710 nm), band 8 (wavelength 800 nm), and band 11 (wavelength 900 nm) based on OIF. Based on these three optimal bands, we further compared the Deeplabv3+ model with U-Net and ENVINet5 model. Our results show that Deeplabv3 + model can capture the potential information in the images at a deeper level with fewer parameters. The MIoU derived from Deeplabv3+, U-Net, and ENVINet5 was 85.54%, 21.21%, and 27.19%, respectively. Also, Deeplabv3 + model solved the problem of sample unbalance. Our results provide a reference for the application of semantic segmentation using multispectral remote sensing data in farmland areas.