Mildew Detection for Stored Wheat using Gas Chromatography-Ion Mobility Spectrometry and Broad Learning Network

Most of the existing methods for wheat mildew detection are biochemical methods, which have the problems of complicated procedures and slow speed. In this paper, a novel wheat mildew detection and classification model is proposed by combining gas chromatography-ion mobility spectrometry (GC-IMS) with a broad learning network (BLN) model. Firstly, the GC-IMS fingerprint spectrums of wheat samples with different degrees of mildew are collected by GC-IMS spectrometer, and then an effective and efficient incremental learning system without the need for deep architecture is constructed to identify these fingerprint spectrums. In the BLN model, ridge regression of the pseudo-inverse is designed to find the desired connection weights, and the new weights can be updated easily by only computing the pseudo-inverse of the corresponding added node. To improve the classification accuracy of the BLN model, incremental learning and the spatial attention mechanism (SAM) are introduced into the model. Experimental results show that the training time of the proposed model is greatly reduced compared to existing deep-learning models. Under the small sample set condition, the mean average accuracy (mAP) of wheat mildew types reaches 90.32%, and the identification precision of early wheat mildew reaches 95.34%. The comprehensive index shows that the neural network model proposed in this paper can be used as an alternative model for deep learning in similar areas of image recognition. The experiment also proved that GC-IMS combined with a broad learning model is an efficient and accurate method for wheat mildew detection.