Classification of tree symbiotic fungi based on hyperspectral imagery and hybrid convolutional neural networks

Abstract

Hyperspectral imagery and machine learning have proven to be powerful, non-invasive, and chemical-free tools for studying tree symbiotic fungi. However, traditional machine learning requires manual feature extraction (feature engineering) of spectral and spatial features of tree symbiotic fungi. Deep convolutional neural networks (CNNs) can extract self and robust features directly from the raw data. In the current study, a deep CNN architecture is proposed to recognize the isolates of dark septate endophytic (DSE) fungal in hyperspectral images. The performance of different CNN approaches (two-dimensional and three-dimensional CNNs) was compared and evaluated based on two independent datasets collected using visible-near-infrared (VNIR) and short-wave-infrared (SWIR) hyperspectral imaging systems. Moreover, the impact of different spectral pre-processing techniques was investigated. The results show that a hybrid CNN architecture (3D-2D CNN), which combines three and two-dimensional CNNs, achieved the best performance for the classification of fungal isolates on SWIR hyperspectral data compared to the same architecture on VNIR hyperspectral data. The best performance is 100% for precision, recall, and overall accuracy. The results also demonstrate that combining different pre-processing techniques on raw SWIR spectra can significantly improve the performance of the CNN models for fungal classification. The hybrid CNN approach with SWIR hyperspectral data provides an efficient method for classifying fungal isolates, which can contribute to the development of accurate and non-destructive tools for evaluating the occurrence of fungal isolates on trees. Such tools can be beneficial for both sustainable agriculture and preserving fungal diversity.