Research On Method And Application Of Deep Forest In Near-infrared Spectroscopy Classification

Near-infrared spectroscopy(NIRs)has the outstanding characteristics of fast analysis,high efficiency,non-destructive and non-polluting.It can be combined with machine learning to achieve qualitative analysis in NIRs detection.However,NIRs data are closely related to the measurement environment,instruments,experimental levels,etc.Samples collected in different batches have significantly different and it is difficult to obtain a large-scale calibration samples,resulting in the shortcomings of insufficient feature learning ability and training difficulties in traditional methods.Its classification result is often poor.Therefore,exploring more effective NIRs classification approaches is the key to improving the efficiency in NIRs tasks.In this paper,the deep forest algorithm is used as the analysis method of NIRs data.It is studied from the two aspects of feature integration learning and representation learning,and constructs the identification model.It is verified in the citrus greening detection and drug fine classification tasks,and the classification performance of NIRs is improved.The main research work of this paper is as follows:(1)Taking the idea of feature integration of deep forest as the research point,this paper proposes a multi-feature fusional cascade forest(FCForest)algorithm.Firstly,the effective features of the original data are extracted,then the cascaded ensemble decision tree is used to fully learn the difference and complementary information between features,and the optimal model structure is adaptively selected.The results showed that under two different NIRs tasks of citrus greening detection and drug identification,the predictive accuracy and model stability of FCForest were significantly better than other mainstream algorithms,and the training time of original deep forest approach was greatly reduced,so it was an effective NIRs identification method.(2)Taking the hierarchical representation of deep forests as the research point,this paper proposes a multi-layered gradient boosting decision trees with adaptive feature(FGBDT)algorithm.By adding an adaptive feature mechanism,feature redundancy and model complexity are greatly reduced,and non-differentiable modules are optimized with target propagation variants to map better spatial distribution of data features.Experiments show that FGBDT can obtain excellent feature distribution representation and has higher accuracy and robustness than other methods.Based on this,cost-sensitive learning is introduced,which shows excellent performance on unbalanced data.Therefore,it is an accurate and reliable NIRs classification method.