??AI for Thrips: Detection in Mango Orchards ??

Thrips, primarily of the genus Thripidae, pose significant agricultural risks, especially in mango orchards.

Thrips damage plants by puncturing leaf and fruit tissues with their mouthparts, siphoning out nutrients, and disrupting photosynthesis and water transport within the plant.

This leads to rust spots, leaf deformation, and reduced chlorophyll, which directly impacts plant health, yield, and fruit quality.

In the figure above:

A. mango tree inflorescences with initial symptoms of floral malformation;

B. highly infested by thrips and black colour (advanced symptom) in the same mango orchard;

C. Frankliniella distinguenda;

D. Frankliniella gardeniae;

E. Frankliniella gemina;

F. Haplothrips gowdeyi.

The threat is amplified in warm, humid environments, which favour rapid thrip population growth and damage expansion.

Today's study, recently published by the team of Chinese researchers, employs a novel remote sensing technique to predict and map thrip damage across mango orchards. Leveraging a Maximum Likelihood Classifier (MLC) and multispectral data, it achieves a robust 91.23% predictive accuracy. The findings offer a foundation for managing thrip outbreaks effectively by integrating the following insights:

Methodology

The study's methodology combines remote sensing and machine learning to predict and analyze thrips damage in mango orchards. Here’s a step-by-step breakdown of the experimental procedure:

1. Selection of Experimental Site

• Location: The experiment was conducted in a mango orchard in Sangongli Village, Sanya City, Hainan Province, China.
• Mango Variety: The orchard grows Guifei mango trees, which are 7-8 years old and feature umbrella-like canopies.
• Tree Specifications: Trees were spaced 4.5 meters apart, with row spacing of 5 meters, ensuring good visibility for remote sensing data collection.

2. Data Collection Setup

The DJI P4 multispectral drone, equipped with six spectral bands (blue, green, red, red-edge, and near-infrared), was selected for image collection. Flights were conducted in September 2023, coinciding with the mango flowering period when thrips activity typically increases. To ensure comprehensive coverage, flight parameters were set with an 80% forward overlap and a 70% side overlap.

Before each flight, a spectral calibration whiteboard was used to calibrate images for lighting and reflectance, ensuring data accuracy across all conditions.

3. Ground Sampling of Mango Trees

• Sampling Points: Twenty trees were randomly selected within the orchard. Each tree had five sampling points (covering different leaf sections).
• Chlorophyll Measurement: A SPAD 502 Plus chlorophyll concentration meter measured chlorophyll content at each sampling point.
• GPS Recording: High-precision GPS recorded sampling point locations to align ground data with aerial imagery in later analyses.

4. Classification of Thrips Damage Levels

Mango leaves were inspected visually, with damage levels categorized based on the extent of rust spots (visible brown streaks on leaves caused by thrips).

Damage Levels: Based on forestry pest assessment standards damage was divided into four levels:

???? healthy (0-10% rust spots),

???? mild (11-20%),

???? moderate (21-50%), and

????severe (51-100%) .

Leaves in each category were recorded, photographed, and used as labelled data for model training and validation.

5. Spectral Data Collection

1?? Hyperspectral Data Collection: Ten mango leaf samples per damage level were taken, each sampled at three points (base, middle, and tip).

2?? Spectral Response Measurement: The ATP9100 hyperspectral spectrometer recorded the reflectance values across wavelengths (300–1100 nm), creating spectral curves for each damage level.

3?? Data Averaging: Reflectance data across damage levels were averaged for each spectral band to reduce noise and enhance accuracy in spectral curve analyses.

6. Selection of Vegetation Indices

• Index Calculation: Fourteen vegetation indices commonly associated with plant health (e.g., NDVI, SAVI, GNDVI) were calculated from the spectral data.
• Correlation Analysis: Pearson correlation analysis determined the relationship between each index and thrip damage severity. Six indices, including NDVI, SAVI, and GNDVI, demonstrated significant correlations with damage levels.
• Feature Index Selection: The Greenness Normalized Difference Vegetation Index (GNDVI) was selected as the primary index for model development due to its high correlation (R2 = 0.82) with thrip damage levels.

7. Model Development Using Maximum Likelihood Classifier (MLC)

The Maximum Likelihood Classifier (MLC), a machine learning algorithm, was chosen for its effectiveness in handling spectral data.

The standard MLC was modified to account for uneven pixel data by adding weighted adjustments to neighboring pixels, improving accuracy.

Using training samples (classified by damage level), a probability density function was created to predict the likelihood of each pixel falling within one of the four damage categories.

8. Validation and Field Verification

A follow-up sampling in January 2024 validated the model’s predictions. Data confirmed high thrip activity in areas predicted by the model.

Field surveys matched the model’s 3D maps, indicating accuracy in predicting damage patterns and supporting pesticide application planning.

Machine Learning Model Testing and Performance

The study compared various classification methods, including SVM, JSR, ASOMP, and MLC. The novel MLC outperformed with an average classification accuracy of 91.23%.

• Classification Accuracy by Damage Level:

1. Severe Damage: 90.17%
2. Moderate Damage: 93.11%
3. Mild Damage: 91.57%
4. Healthy: 90.05%

Inversion Prediction and Thrip Damage Distribution

• 3D Heatmap Visualization: A 3D map revealed thrip damage concentrated in specific orchard zones, notably areas with abundant young leaves which attract thrips.
• Thrip Damage Over Time:

- September 2023: Mild thrip damage predominated (37.9% of affected area), with severe damage in 10.5%.

- January 2024: Severe damage increased to 30.8% due to favourable climatic conditions, especially dry monsoon winds.

?? Conclusions

The predictive model aids in timely, large-scale orchard monitoring, which is vital for pest management and pesticide application.

Future enhancements, such as integrating additional spectral bands or expanding UAV applications for broader orchard monitoring, are recommended for even greater accuracy.        

This method provides a dynamic and efficient solution for thrip damage monitoring, allowing mango producers to mitigate damage effectively and sustainably manage orchard health.

?? Citation

Wang, L.; Tang, Y.; Liu, Z.; Zheng, M.; Shi, W.; Li, J.; He, X. Prediction of Thrips Damage Distribution in Mango Orchards Using a Novel Maximum Likelihood Classifier. Agronomy 2024, 14, 795. https://doi.org/10.3390/agronomy14040795

?? Useful reading on AI against Thrips

The articles in this selection are protected by publisher's paywall

?? What's on

If you'd like to receive the regular 'AI in Agriculture' newsletter in your inbox, simply add your email to my mailing list.

Join over 9,100 readers who enjoy weekly updates on AI advancements in agriculture!

AI for Fruits Phenotyping and Quality Assurance

To get more information about fruit phenotyping capabilities with mobile - ask Petiole Pro