基于深度学习的无人机多光谱图像柑橘树冠分割

Deep Learning-based Segmentation of Citrus Tree Canopy from UAV Multispectral Images

树冠信息的准确获取是研究柑橘树生长及产量预测的重要指标,但复杂的树木结构给树冠的准确提取带来一定影响,深度学习的快速发展为柑橘树冠信息准确获取提供了可能。采用一种新的基于深度学习的柑橘树冠自动提取方法,即将消费级无人机采集的多光谱图像与一种新的深度学习模型U2-Net结合,通过对获取的图像进行几何变换以构建柑橘树冠图像数据集。将U2-Net模型和当前3种主流深度学习模型(即PSPNet、U-Net和DeepLabv3+)分别在3个典型试验分区进行试验以提取柑橘树冠信息,并对提取结果进行对比分析。结果表明,在3个试验分区,U2-Net模型的柑橘树冠提取精度最高,其中交并比(IoU)、总体精度(OA)和F1分数(F1-score)分别为91.93%、92.34%和93.92%。与其他3种深度学习模型相比,U2-Net模型的IoU、OA和F1-socre分别提高了3.63%~8.31%、1.17%~5.25%和1.97%~4.91%。此外,U2-Net模型柑橘树冠提取面积和测量面积之间具有较高的一致性,3个试验分区决定系数(R2)均高于0.93,且与其他3种深度学习模型相比,U2-Net模型的错误率也较低,均方根误差(RMSE)为1.35 m2,均方误差(MRE)为8.15%。此研究将无人机多光谱图像与U2-Net模型相结合的方法能够实现柑橘树冠的精确提取,且提取树冠轮廓完整性较好,可为柑橘动态生长变化监测和产量预测提供基础数据与技术支撑。

Accurate acquisition of canopy information was an important indicator to study the growth and yield prediction of citrus trees,but the complex tree structure had a certain impact on the accurate extraction of canopy.The rapid development of deep learning provided the possibility for accurate acquisition of citrus canopy information.Based on this,an automatic extraction method of citrus canopy based on deep learning was adopted,which combined the multispectral images collected by consumer drones with a new deep learning model U 2-Net,and constructed a dataset of citrus tree canopy by performing geometric transformation on the acquired images.The U 2-Net model and the current three mainstream deep learning models(i.e.PSPNet,U-Net and DeepLabv3+)were tested in three typical sample regions to extract the citrus canopy,and the extracted results were compared.Results showed that:in the three sample regions,the U 2-Net model had the highest extraction accuracy of citrus canopy,with the intersection over union(IoU),overall accuracy(OA)and F1-score of 91.93%,92.34%and 93.92%,respectively.Compared with the other three deep learning models,the IoU,OA and F1-socre of the U 2-Net model were improved by 3.63%-8.31%,1.17%-5.25%and 1.97%-4.91%,respectively.In addition,the U 2-Net model had high consistency between the extraction area and the measured area of the citrus canopy,and the coefficient of determination(R 2)of the three sample regions was higher than 0.93,and the error rate of the U 2-Net model was lower than that of the other three deep learning models,with a root mean square error(RMSE)of 1.35 m 2 and a mean relative error(MRE)of 8.15%.The results showed that the method of combining the U 2-Net model with the multi-spectral image of the drones can realize the accurate extraction of citrus canopy,and the extracted canopy contour had a good integrity,which can provide basic data and technical support for the monitoring of citrus dynamic growth changes and yield prediction.