NIR hyperspectral imaging with machine learning to detect and classify codling moth infestation in apples

Nondestructive Detection of Codling Moth Infestation in Apples Using Pixel-Based NIR Hyperspectral Imaging with Machine Learning and Feature Selection

“…After the acquisition and correction of the hyperspectral images, the spectral information of the infested and healthy tissue was automatically extracted from ROIs using the algorithm described in Figure 2 . Since the CM larvae, especially the first generation, mostly enter apples from the calyx end [ 47 ] and the initial results by Ekramirad et al [ 30 ] showed that the highest infestation classification accuracy achieved in images from the calyx view, the ROI to extract infested pixels was segmented around the calyx end. This novel method can select the complete infested region with pixels in the healthy region as few as possible to obtain a precise infested region for subsequent classification.…”

“…HSI has been investigated as a rapid and relatively low-cost nondestructive technique in the quality assessment of apples. This application mainly falls into three categories including external quality, internal quality, and pest detection [ 29 , 30 ]. Regarding external quality of apples, HSI was used to evaluate defects (e.g., surface defects and bruising) because of its ability to penetrate beneath the apple’s skin.…”

“…HSI also has been used for nondestructive prediction of internal quality of apples such as the nutritional value, texture, and flavor components, and in estimating physiochemical parameters such as vitamin and sugar content [ 33 ]. Additionally, HSI has been tested for safety assurance of apples through effective detection of pests [ 30 , 34 ]. In 2021, Ekramirad et al [ 30 ] determined the best classification result for CM detection in apples using NIR HSI and the mean spectra extraction method on dataset consisting of three apple orientations (clayx, sides and bottom).…”

“…Additionally, HSI has been tested for safety assurance of apples through effective detection of pests [ 30 , 34 ]. In 2021, Ekramirad et al [ 30 ] determined the best classification result for CM detection in apples using NIR HSI and the mean spectra extraction method on dataset consisting of three apple orientations (clayx, sides and bottom). Applying partial least squares-discriminant analysis (PLS-DA) classifier, they obtained an overall validation accuracy of 81.04%.…”

“…However, they reported that by reducing the data dimensions using the sequential forward selection method, their classification accuracy was enhanced to 82%. In the same study conducted by Ekramirad et al [ 30 ], they applied a second pixel-wise method instead of the mean spectra extraction and found an accuracy as high as 98.2% in classifying the infested and healthy pixels using the random forest (RF) classifier. However, they had manually segmented a rectangular ROI around the calyx end of the apple sample to extract pixels data by spectra for infested and healthy, which is a cumbersome subjective task that increases the processing time, hence there is a need to develop a new method for automatic extraction of target pixels.…”