AI-Powered Early Diagnosis of Strawberry Wilt

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 4 2024

In a paper published in the journal Precision Agriculture, researchers investigated using hyperspectral imagery in conjunction with machine learning (ML) algorithms to detect and categorize strawberry infections with early Fusarium wilt by developing six ML models: artificial neural network (ANN), decision tree (DT), K-nearest neighbor (KNN), support vector machine (SVM), multinomial logistic regression (MLR), and Naïve Bayes (NB).

The ANN model achieved the highest accuracy at accurately predicting physiological stress indicators like stomatal conductance and photosynthesis even before visual symptoms appeared. This approach significantly enhanced early disease detection and management.

Background

Past work has shown that strawberries, a leading global crop, are significantly impacted by Fusarium oxysporum f sp. fragrariae (Fof), a dangerous soilborne pathogen. Research has focused on integrating hyperspectral imaging and ML to detect early infections and physiological stress in strawberries before visual symptoms appear. Studies demonstrated high accuracy in disease detection using hyperspectral data and ML models, improving early diagnosis and management strategies.

Experimental Procedure Overview

The study was conducted at the Higher Technical School of Agricultural Engineering (ETSIA), University of Seville, using strawberry plants from the cv. Palmeritas. These plants were grown in 1-liter pots arranged in a completely randomized block design with four blocks and seven treatments.

Six treatments—F74, F111, F115, F134, F138, and F141—involved distinct Fusarium oxysporum strains, with one treatment acting as a control (non-inoculated). The plants were irrigated twice weekly with two fertilizer solutions while housed in a growth room with regulated temperatures and a twelve-hour photoperiod.

Fungal isolates were stored at -80°C and cultivated on an agar malt asparagine peptone (AMAP) culture medium for seven days. Strawberry plants were inoculated by root-dipping in these suspensions one week before taking measurements. The F74 isolate, known for causing Fusarium wilt with a distinct genotype and phenotype, was compared with other non-pathogenic isolates.

Hyperspectral imaging was performed using a headwall nano-hyperspec camera with a spectral range of 397 to 1003 nm and a resolution of 272 bands. Researchers fixed the camera on an aluminum frame and placed the pots on a conveyor belt. Measurements were taken weekly for four weeks, with controlled lighting provided by tungsten halogen lamps. Data were processed using Python to correct for spectral reflectance, smooth images, and normalize the data.

The team calibrated hyperspectral data to extract leaf tissue signatures, with simultaneous measurements of disease severity and leaf gas exchange. Six machine learning (ML) models—ANN, DT, K-NN, SVM, MLR, and NB—were used for regression and classification. Analysts used mean absolute error (MAE), root mean square error (RMSE), F1 Scores, and analysis of variance (ANOVA) to evaluate the performance of the models.

Wilt Assessment

On the final sampling date (M4), significant differences in Fusarium wilt severity were observed between control and inoculated plants. Only two isolates, F74 and F141, showed noticeable disease symptoms. F74 demonstrated substantial differences from the control, while F141 showed high symptom variability, limiting its statistical significance.

Leaf gas exchange data revealed that all the Fusarium oxysporum isolates significantly reduced the stomatal conductance (g_s) and the leaf photosynthetic rate (A) relative to the control group. F74 had the most severe impact, causing a 35% reduction in g_s at the first sampling date, which worsened to 90% by the last two sampling dates. Similarly, F74 caused up to a 90% reduction in A by the end of the study. Other isolates also reduced g_s and A, though less severely than F74, with significant reductions observed from the third sampling date onwards.

ML models, including ANN, DT, and others, were employed to classify and estimate Fusarium wilt severity based on physiological parameters. The ANN model achieved the highest classification accuracy, correctly identifying 100% of healthy plants and performing well in categorizing plants based on g_s and A. The model demonstrated strong regression performance with R² values of 0.84 for g_s and 0.81 for A and low RMSE and MAE values, indicating its effectiveness in predicting physiological parameters of both inoculated and non-inoculated plants.

The spectral signatures of the plants were recorded during the investigation using hyperspectral imaging, which showed notable variations in reflectance between healthy and damaged plants. While variations were less obvious in the near-infrared area (700–900 nm), the green to red spectral bands (550–670 nm) showed the most dramatic diversity in reflectance.

The ANN model was particularly effective in utilizing these spectral signatures for classification, achieving high accuracy in identifying the severity of physiological damage. This approach underscores the potential of hyperspectral imaging combined with ML to assess plant health and disease severity in agricultural research.

Conclusion

To sum up, pathogenic Fusarium oxysporum isolates caused rapid wilting of strawberry plants and severe reductions in leaf gas exchange variables. Non-pathogenic isolates also resulted in some reduction, although less pronounced. The combined use of ML models and hyperspectral imaging effectively classified healthy and diseased plants based on physiological impairment, with the ANN model achieving over 80% accuracy. Future improvements were needed to enhance model performance across varying stress conditions and plant cultivars.