基于卷积神经网络遥感图像的荔枝花期长势评估

文飞; 莫嘉维; 胡宇琦; 兰玉彬; 陈欣; 陆健强; 邓小玲

doi:10.7671/j.issn.1001-411X.202203040

基于卷积神经网络遥感图像的荔枝花期长势评估

The assessment of litchi flowering growth based on remote sensing image of convolutional neural network

摘要

摘要:
目的通过无人机获取荔枝冠层的遥感图像，评估每棵荔枝的开花率，以期为后续荔枝花期疏花保果、精准施肥施药提供决策依据。

方法以遥感图像为研究对象，利用实例分割的方法分割每棵荔枝冠层后，结合园艺专家的综合判断，按开花率为0、10%~20%、50%~60%、80%及以上将开花率分为4类，使用ResNet、ResNeXt、ShuffleNetv2进行开花率分类比较，试验过程中发现ShuffleNetv2在识别准确率、参数量、训练和验证时间都有很大优势；在ShuffleNetv2上引入了空间注意力模块(Spatial attention module，SAM)后，增加了模型对位置信息的学习，在不显著增加参数量的情况下，提升荔枝冠层花期分类的精度。

结果通过对多个主流深度神经网络的比较分析，ResNet50、ResNeXt50、ShuffleNetv2的分类精度分别达到85.96%、87.01%和86.84%，而改进后的ShuffleNetv2分类精度更高，达到88.60%；ResNet50、ResNeXt50、ShuffleNetv2和改进后的ShuffleNetv2对测试集单张冠层图像验证的时间分别为8.802、9.116、7.529和7.507 ms，改进后的ShuffleNetv2单张冠层图像验证时间最短。

结论改进后的ShuffleNetv2能够挖掘学习更为细节的荔枝冠层花期信息，具有较高的识别准确率，对荔枝花期的评估有很大的优势，可为荔枝保花疏花、生产精准管控提供智能决策支持。

Abstract:
Objective In order to provide decision-making basis for subsequent litchi flower thinning, fruit retention and precise fertilization application, this work evaluated the flowering rate of each litchi by analyzing UAV remote sensing images of litchi canopy.

Method The remote sensing image of each litchi canopy was segmented through instance segmentation algorithm. The flowering rates were classified into four categories combining with comprehensive judgment of horticultural experts, which were 0, 10%−20%, 50%−60%, 80% and above. ResNet50, ResNeXt50 and ShuffleNetv2 were adopted to compare flowering rate classification. Due to the great advantages in recognition accuracy, number of parameters, training and verification time, ShuffleNetv2 was adopted as the instance segmentation algorithm, and optimized by introducing the spatial attention module (SAM) to increase the model’s learning of location information, and improve the accuracy of litchi canopy flowering classification without significantly increasing the number of parameters.

Result Through comparison of the mainstream deep learning algorithms, the classification accuracy of ResNet50, ResNeXt50 and ShuffleNetv2 reached 85.96%, 87.01% and 86.84% respectively, and the improved ShuffleNetv2 reached 88.60%, higher than the above three algorithms. The verification time of single canopy image on test set using ResNet50, ResNeXt50, ShuffleNetv2 and the improved ShuffleNetv2 were 8.802, 9.116, 7.529 and 7.507 ms respectively, showing that the improved ShuffleNetv2 single canopy image got the shortest verification time.

Conclusion The improved ShuffleNetv2 can excavate and learn more detailed flowering information of litchi canopy, with high recognition accuracy and great advantages in the evaluation of litchi flowering, providing an intelligent decision support for flower protection and sparseness, and precise control of production.

HTML全文

参考文献(28)

施引文献

资源附件(0)