基于CTGAN与GDMPA-RF算法的活立木含水率诊断方法优化研究

Optimization of Living Trees Moisture Content Diagnosis Method Based on CTGAN and GDMPA-RF Algorithm

活立木含水率的精准实时诊断是智慧林业领域的重要研究内容,其可为植物生理状态分析、林区生态水文调控、林火预警防范等做出关键指征。基于无线声发射传感器网络(Wireless Acoustic Sensor Network,WASN)系统的含水率测定方法既可实现高效无损探测,又能长期野外部署,尤为适合林场实际需求。为了进一步提升WASN的辨识准确率,首先利用条件表格生成对抗网络(Conditional Tabular GAN,CTGAN)对所采集的AE特征进行数据增广,其次基于分布式梯度提升框架(Light Gradient Boosting Machine,LightGBM)对扩增后的混合数据集进行特征优选,然后提出了黄金正弦动态海洋捕食者算法优化的随机森林(Golden-Sine Dynamic Marine Predators Algorithm-Random Forests,GDMPA-RF)策略,并以此建立含水率精准反演模型。实验对比结果显示,基于优选特征子集构建的GDMPA-RF模型在立木含水率诊断性能强化方面效果最佳,其准确率(Accuracy)、精确率(Precision)、F1分数(F1-Score)、加权平均(Weighted Average)和AUC分别为99.17%、99.52%、98.14%、0.9943和0.9850,均高于鲸鱼优化算法等结合RF模型的评估指标,说明方法具有优良的监测效能,较好地优化了活立木树干含水率的在线实时推演精度。

Accurate and real-time diagnosis of living trees moisture content(MC)is an important research issue in the field of smart for-est,which can provide key indications for plant physiological state analysis,eco-hydrological control of forest areas and forest fire pre-vention.Wireless acoustic emission sensors(WASN)based MC diagnosis method is particularly suitable for forest as they are highly effi-cient and non-destructive,and can be deployed in the field for long periods of time.In order to further improve the recognition accuracy of WASN.Firstly,conditional tabular generative adversarial network(CTGAN)is used for data augmentation of the collected AE fea-tures,secondly,feature optimization is performed on the augmented hybrid dataset based on distributed gradient boosting framework(LightGBM),and then a novel golden-sine dynamic marine predators algorithm-random forests(GDMPA-RF)strategy is proposed to es-tablish an accurate inversion model of MC.The experimental comparison results show that the GDMPA-RF model based on the preferred feature subset is the most effective in enhancing the diagnostic performance of standing wood MC,with the Accuracy,Precision,F1-Score,Weighted Average and AUC of 99.17%,99.52%,98.14%,0.9943 and 0.9850 respectively,all of which are higher than the e-valuation indices of other optimized algorithm combined with RF,such as whale optimization algorithm,indicating that the method has excellent monitoring efficacy and greatly optimizes the accuracy of online real-time detection of moisture content of live tree trunks.