**1. Introduction**

Crop diseases greatly impact the yield and quality of agricultural products, as they can easily cause stem and leaf death, thereby leading to plant decay [1]. In this way, such diseases affect human food security and food safety. Therefore, research on technologies for crop disease diagnosis is of great significance for the early warning and control of these diseases. The traditional diagnosis method used for crop diseases mainly relies on manual diagnosis, which is based on the experience of the examiner. Although this method is simple and convenient, it consumes a great deal of manpower and allows for a high degree of subjectivity, which can lead to misdiagnosis. Currently, the most objective and accurate disease detection methods available are based on laboratory biochemical tests (e.g., the polymerase chain reaction (PCR), nucleic acid hybridization, and DNA microarray techniques) [2–4]. Although laboratory-based biochemical detection methods feature the advantage of high identification accuracy, their involved sampling and detection steps require professional operation, are associated with high costs, are lengthy to conduct, and are difficult to conduct on a large-scale [5,6]. In recent years, the rapid development of machine vision and spectral imaging technologies has enabled the quick

**Citation:** Zhang, X.; Wang, Y.; Zhou, Z.; Zhang, Y.; Wang, X. Detection Method for Tomato Leaf Mildew Based on Hyperspectral Fusion Terahertz Technology. *Foods* **2023**, *12*, 535. https://doi.org/10.3390/ foods12030535

Academic Editor: Ana Teresa Sanches-Silva

Received: 1 December 2022 Revised: 14 January 2023 Accepted: 19 January 2023 Published: 25 January 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

detection of crop diseases. Such technologies include visible/near-infrared imaging, multispectral/hyperspectral imaging, and chlorophyll fluorescence imaging, which have all been applied to crop disease detection [7,8]. Although this represents progress, most existing studies only discriminate the grade of crop disease by the reflective properties and apparent outer characteristics of the diseased leaves. Because the internal damage of diseased leaves cannot be detected, it remains difficult to achieve the combined analysis of internal and external damage caused by fungal diseases.

In recent years, hyperspectral technology has attracted increasing research interest in the context of disease detection, owing to its merits of featuring high-resolution and integrated mapping. Spectral imaging technology can obtain the spectral image data cubes of the tested sample, thereby accurately obtaining the image information and spectral reflection intensity distribution characteristics of each test sample in each waveband. Fazari et al. [9] established a three-dimensional CNN model using hyperspectral imaging to classify olive anthrax, which performed with a prediction accuracy of 95.73%. Zhang et al. [10] used visible light imaging on downy mildew in combination with machine learning methods to quickly and accurately estimate the severity of cucumber downy mildew in a greenhouse. Image features that had a high correlation with the actual value of greenhouse cucumber downy mildew severity were then used to construct a shallow machine-learning estimation model. The results showed that there was a good linear relationship between the severity of greenhouse cucumber downy mildew estimated by the model and the actual value. Qin et al. [11] proposed a feature band extraction method combining an improved competitive adaptive reweighting algorithm (CARS) and a successive projections algorithm (SPA) with disease information to establish an early detection model of cucumber downy mildew. With this model, the difficult problem of conducting the early detection of cucumber downy mildew was solved.

Terahertz (THz) radiation refers to long wavelength electromagnetic waves with a frequency range of 0.1–10 THz (corresponding to wavelengths of 30 μm–3 mm). THz waves penetrate deeply into the medium and their high correlation helps to determine the exact refractive index and absorption coefficient of a given sample. THz spectroscopy can be utilized to analyze macromolecules and components inside of crops due to the transmission properties of the radiation, which gives it unique advantages in the application of biological information detection. Some researchers have carried out a preliminary attempt at the THzbased detection of crops and agricultural products [12,13]. Di Girolamo et al. [14] imaged 50 chestnuts that were partially infected with Pygmy fungus in the low THz frequency range by means of a homemade 0–0.1 THz small portable imaging system. By assuming different moisture densities and different physical structures of healthy and unhealthy chestnuts, the relationship between the physical parameters (mass or volume) of chestnuts and the light attenuation of healthy and infected chestnuts was tentatively resolved. The results showed that the index of light attenuation combined with the measurement of chestnut weight or volume could successfully identify whether a given chestnut was healthy or diseased. Li et al. [15] employed a recognition model based on a THz spectroscopy technique to analyze data for apple ring rot and cucumber powdery mildew. The researchers established recognition models for common crop diseases based on K-nearest neighbor, SVM, and BP neural network algorithms, respectively, with a correlation coefficient Rp of 0.9649. Their findings demonstrated that hyperspectral and THz technology could be used to detect crop diseases. However, it remains difficult to obtain the internal and external indicators of crop diseases from either external characterization or by using only a single method, and the prediction accuracy also needs to be further improved.

Tomato leaf mold, also known as black mold and black hair, is a tomato disease caused by *Fulvia fulva* (Cooke) Cif. Tomato leaf mildew mainly affects the leaves of infected plants, and in severe cases, also affects the stems, flowers, and fruits. In the early stages of the disease, yellow-green spots with obscure edges appear on the front of affected leaves, while a grayish-white mildew layer appears on the back of the leaves. When the humidity is high, the leaf surface lesions can also grow a mildew layer. After the conidia of tomato leaf mold

invade the tomato leaves, they cause changes in the sugars, lipids, proteins, and nucleic acids inside of the leaves. Existing crop disease detection models employ only a single detection method, and such existing methods are unable to fully reflect the condition of the diseased crops. Therefore, this study acquired the near-infrared hyperspectral data, THz power spectrum, and absorbance time-domain spectral data of tomato leaf mildew samples from different infection grades, and carried out a study on a detection model combining both internal and external features of tomato leaf mildew. Through the spectral analysis of tomato leaves under different characteristic frequency bands, a high-precision prediction model of tomato leaf mildew was established.
