نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه گیاهپزشکی، دانشکده تولید گیاهی، دانشگاه علوم کشاورزی و منابع طبیعی گرگان، گرگان، ایران.
2 گروه گیاهپزشکی، دانشکده تولید گیاهی، دانشگاه علوم کشاورزی و منابع طبیعی گرگان، گرگان، ایران.
3 گروه مهندسی مکانیک بیوسیستم، دانشکده مهندسی آ ب و خاک، دانشگاه علوم کشاورزی و منابع طبیعی گرگان، گرگان، ایران
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Although previous studies have reported spectral signatures of some common fungal diseases of tomato for their early detection using hyperspectral image processing methods, these studies are generally limited to identifying the occurrence or absence of the disease and do not detect the severity of disease progression. Therefore, in this study, an attempt has been made to analyze RGB and hyperspectral images of plant leaf samples, in addition to the early detection of tomato leaf spot disease caused by Alternaria alternata, to predict the time elapsed since inoculation of the fungal spore suspension as the rate of disease progression. Visible and hyperspectral data were collected from leaves with two-day intervals for 18 days after inoculation and before the appearance of visible symptoms of the disease. The results showed that by extracting features from the disease's hyperspectral signature identified in previous research (450, 550, 650, and 900 nm), the random forest method optimized with the Harris hawk’s algorithm can detect the occurrence of the disease only six days after inoculation with an accuracy, precision, and recall of 97, 92, and 94 %, respectively. Also, the random forest optimized with the particle swarm optimization was able to correctly predict the disease progression in the first eight days after inoculation with a coefficient of determination and mean squared error of 0.91 and 0.011, respectively. Compared to the performance of hyperspectral data, the data obtained from the analysis of visible images showed a much weaker performance. The findings of this study showed that hyperspectral image analysis can be a powerful tool in the early detection and prognosis of plant diseases.
کلیدواژه‌ها [English]
Early blight caused by Alternaria alternata, is one of the most prevalent and destructive fungal disease in tomato plants. Timely and accurate detection is critical for minimizing yield losses. However, conventional and molecular diagnostic methods are often time-consuming, costly, or require visible symptom development. In recent years, hyperspectral imaging (HSI) has emerged as a powerful, non-destructive technique for early detection of plant stresses. Numerous studies have demonstrated the high accuracy of this method in detecting fungal diseases, including early blight. However, most of these investigations have primarily focused on the binary classification of disease presence or absence. The present study aimed to determine the occurrence and severity (prognosis) of the disease caused by a fungal pathogen, using RGB and hyperspectral imaging in tomato plants. The primary objective was to develop a robust and accurate predictive model by applying machine learning algorithms to features extracted from the image data, aiming to assess disease severity with high performance and minimal error.
Tomato seedlings were grown under controlled greenhouse conditions and artificially inoculated with spore suspension of Alternaria alternata. From day 2 to day 18 after inoculation, hyperspectral and RGB images were captured from healthy and infected leaves at two-day intervals. Image data were pre-processed by background removal, disease region segmentation, and extraction of color, textural, and morphological features. Four key hyperspectral bands (450, 550, 650, and 900 nm), previously identified as effective for early disease detection, were selected for spectral signature. Three machine learning models—artificial neural network (ANN), random forest (RF), and support vector machine (SVM) were employed, and their hyperparameters were optimized using four metaheuristic algorithms: genetic algorithm, particle swarm optimization, whale optimization algorithm, and Harris hawks optimization. The objective was to maximize classification accuracy of disease presence and minimize prediction error for disease progression severity. Models were evaluated using five-fold cross-validation, and their performance was assessed based on accuracy, precision, and recall (in the detection of the occurrence of the disease), as well as mean squared error (MSE) and coefficient of determination (R²) (in the detection of severity of the disease).
The results showed that by extracting features from the disease's hyperspectral signature identified in previous research (450, 550, 650, and 900 nm), the random forest method optimized with the Harris hawks algorithm can detect the occurrence of the disease only six days after inoculation with an accuracy, precision, and recall of 97, 92, and 94 %, respectively. Also, the random forest optimized with the particle swarm optimization was able to correctly predict the disease progression in the first eight days after inoculation with a coefficient of determination and mean squared error of 0.91 and 0.011, respectively. Compared to the performance of hyperspectral data, the data obtained from the analysis of RGB images showed a much weaker performance. RGB features were less sensitive at earlier stages of infection, whereas hyperspectral data enabled detection even before visible symptoms appeared. The detection accuracy with RGB images remained only around 70% even 18 days after disease transmission.
This study underscores the potential of integrating hyperspectral imaging with machine learning for the early and accurate detection of fungal diseases in tomato plants. In comparison with hyperspectral imaging, the outcomes of RGB image analysis exhibited markedly lower performance, emphasizing the critical role of hyperspectral data in the precise detection and monitoring of plant diseases. Based on the findings, the random forest model, optimized using the Harris hawks algorithm, demonstrated high efficiency in the early and rapid detection of the disease, potentially contributing significantly to informed management decisions and timely agricultural interventions.
All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.
Data available on request from the authors.
The authors gratefully acknowledge the support provided by the Gorgan University of Agricultural Sciences and Natural Resources, which contributed significantly to the completion of this study.
The authors avoided data fabrication, falsification, plagiarism, and misconduct.
The author declares no conflict of interest.
)