基于无人机多光谱影像和关键点检测的雪茄烟株数提取

为从无人机遥感影像中准确识别烟草,实现植株定位与计数,以雪茄烟草植株为研究对象,提出一种新的深度学习模型。区别于传统的利用检测框识别目标,本文模型利用少量的关键点学习烟草中心形态学特征,并采用轻量级的编、解码器从无人机遥感影像快速识别烟草并定位计数。首先,提出的模型针对烟草植物形态学特点,通过中心关键点标注的方法,使用高斯函数生成概率密度图,引入更多监督信息。其次,对比不同主干网络在模型中的效果,ResNet18作为主干网络时平均精度大于99.5%,精度和置信度都高于测试的其他主干网络。而MobileNetV2在CPU环境下达到运行效率最优,但平均置信度相对较低。使用损失函数Focal Loss与MSE Loss结合的Union Loss时,平均精度大于99.5%。最后,利用不同波段组合作为训练数据,对比结果发现使用红边波段更有助于模型快速收敛且能够很好地区分烟草和杂草。由于红边波段与植株冠层结构相关,使用红边、红、绿波段时平均精度达到99.6%。本文提出的深度学习模型能够准确地检测无人机遥感影像中的烟草,可为烟草的农情监测提供数据支持。

Counting Cigar Tobacco Plants from UAV Multispectral Images via Key Points Detection Approach

Tobacco is an important industrial crop in China. The survival rate and growing status of tobacco plants after being transplanted to the field are essential for the field management and yield predictions. However, counting the number of live plants is traditionally conducted by labors, which is time consuming and expensive. Unmanned aerial vehicle is a cost-effective option for monitoring croplands and plantations. However, visual inspection for such images can be a challenging and biased task, specifically for locating and detecting plants. As tobacco plant has a characteristic center-oriented feature, a novel deep-learning algorithm was developed to locate and count tobacco plants via key points detection method, instead of using a common bounding-box object-detection approach. The proposed deep learning algorithm was tested on the cigar plants. In the algorithm, the center of each plant was firstly annotated with a point, and a Gaussian probability density was derived to provide useful information of morphological features. Secondly, different backbones and loss functions in the proposed algorithm were evaluated. Using ResNet18 as a backbone provided the most accurate prediction of the plant number (average precision higher than 99.5%). MobileNetV2 was the most efficient backbone, but the uncertainty of predictions was higher than that of ResNet18. The combination of Focal Loss function and MSE Loss function (Union loss) reached the highest accuracy (average precision higher than 99.5%) while reduced the uncertainty. Finally, the evaluation of different combinations of multispectral bands showed that the combination of red-edge, red, and green bands had a better performance than using red, green, and blue bands in differentiating tobacco plants and weeds, resulting in less uncertainty in the tobacco plant detection. The proposed algorithm can accurately locate and count tobacco plant in the UAV images, providing an effective tool and a valuable data support for planting high-quality tobaccos.