?? AI for Citrus Orchards: Drones for Yield Prediction ??

Yield prediction is the process of estimating the expected production of a crop based on various factors such as plant health, weather conditions, and historical data.

Yield prediction with artificial intelligence (using computer vision, deep learning and other technologies) involves analyzing high-resolution images and data to accurately forecast crop production by identifying patterns and anomalies that human observation might miss.

Smartphone-based yield prediction leverages mobile device camera and AI algorithms to capture and analyze crop images in the field, providing farmers with immediate, on-the-spot yield estimates.

Important Announcement

If you wish to get regular "AI in Agriculture" newsletter to your mailbox, please, add your email to my mailing list.

Join more than 7,400 readers who get weekly emails about AI in Agriculture

Unmanned Aerial Vehicles (UAVs) are one of the most helpful tools in citrus orchards because they automate the collection of high-resolution imagery and data, enabling precise monitoring, yield estimation, and efficient management of large-scale orchards. We've already have an article about disease detection with drones for citrus crops. Today we'll go further and look at yield prediction.

But firstly, as usual, let's clarify what types of UAVs are available for agriculture.

The most widely used types of UAVs in agriculture are quadcopters and fixed-wing UAVs.

On the photo above different types of UAVs are represented: (a) Fixed Wing, (b) Single rotor, (c) Quadcopter (d) Hexacopter, and (e) Octocopter. Source: Hanif et al., 2022

On the photo above: (a) Fixed-wing (eBee?); (b) helicopter (Hornet Maxi); (c) octocopter; (d) blimp; (e) flapping-wing (SmartBird); and (f) parafoil-wing (Tetracam). Source: Tsouros et al., 2019

Enhancing citrus fruit yield investigations through flight height optimization with UAV imaging

Country: ???? South Korea

Published: 2 July 2024

This study investigates the optimal flight altitude for UAV imaging to accurately estimate citrus fruit yield, aiming to improve efficiency and accuracy in agricultural management.

The researchers conducted field studies in Jeju Island, South Korea, capturing images of Citrus unshiu trees at five different altitudes (30 m, 50 m, 70 m, 90 m, and 110 m) using UAVs equipped with RGB cameras. The images were processed using histogram equalization and various vegetation indices, such as IPCA and CIVE, to enhance fruit detection. Statistical analyses, including t-tests and Wilcoxon rank-sum tests, were applied to evaluate the accuracy of fruit yield estimates at different altitudes.

The study found that the 30 m altitude provided the most accurate yield estimates, with histogram-equalized images detecting up to 10% more fruit compared to untreated images.

The IPCA vegetation index also produced yield estimates close to actual counts, while higher altitudes showed decreased accuracy. The findings highlight the importance of optimizing flight altitude and image processing techniques for accurate yield prediction, potentially reducing resource waste and labour in agricultural practices.

These results can be practically applied by agricultural researchers and farm managers to enhance precision in crop yield estimation and resource management.

Technologies/methods used

• UAVs with RGB cameras
• Histogram equalization
• IPCA and CIVE vegetation indices
• MATLAB for image processing
• Microsoft Excel and R for statistical analysis

For further details, refer to the research article by Kwon et al. (2024) .

Photo above: Image of each height (30?m, 50?m, 70?m, 90?m, 110?m) of C. unshiu trees. (A) Untreated RGB image; (B) histogram equalized RGB image. Source: Kwon et al., 2024

Photo above: image of each height (30?m, 50?m, 70?m, 90?m, 110?m) of C. unshiu fruit detected from trees. (A) Fruit from untreated RGB image; (B) fruit from histogram equalized RGB image. Source: Kwon et al., 2024

Source: Kwon et al., 2024

Tree-level citrus yield prediction utilizing ground and aerial machine vision and machine learning

Country: ???? United States

Published: 9 June 2022

This study developed and compared three machine learning models for predicting citrus yield at the tree level, utilizing a combination of UAV imagery and ground-based fruit detection to enhance accuracy.

The research involved three models:

1. Model-1 used only UAV multispectral data and tree structural parameters;
2. Model-2 incorporated UAV data and ground-based fruit counts from images taken from one side of the tree;
3. Model-3 added fruit counts from both sides of the tree.

The models were trained using four machine learning algorithms: Gradient Boosting Regression (GBR), Random Forest Regression (RFR), Linear Regression (LR), and Partial Least Squares Regression (PLSR). A total of 48 citrus trees were used for data collection and model validation, with the accuracy of yield predictions assessed through the Mean Absolute Percentage Error (MAPE).

The study found that Model-2, which combined UAV data with single-side fruit counts, provided the most accurate yield predictions with a MAPE of 23.45%. This model outperformed Model-1, which had a MAPE of 35.59%, and was comparable to Model-3, which had a MAPE of 25.72%.

The results indicate that integrating UAV imagery with ground-based fruit detection significantly improves yield estimation accuracy, making Model-2 the preferred method due to its balance of accuracy and data collection simplicity.

These findings are particularly beneficial for citrus growers and researchers focused on improving yield prediction accuracy and optimizing resource allocation.

Technologies/methods used

• UAVs with multispectral and RGB cameras
• Agroview cloud-based platform for data processing
• YOLOv3 object detection algorithm
• Gradient Boosting Regression (GBR)
• Random Forest Regression (RFR)
• Linear Regression (LR)
• Partial Least Squares Regression (PLSR)

For further details, refer to the research article by Vijayakumar et al. (2022) .

A boxplot is a simple chart that shows the range and spread of a set of data, highlighting where most of the values lie. It helps you quickly see the minimum, maximum, and the middle range of the data, making it easier to spot any unusual values or patterns.

Intelligent Integrated System for Fruit Detection Using Multi-UAV Imaging and Deep Learning

Country: ???? Ukraine, ???? Poland

Published: 16 March 2024

This study introduces an advanced system that integrates multi-UAV imaging with deep learning techniques to enhance the accuracy and efficiency of fruit detection and counting in orchards.

The researchers utilized a fleet of UAVs equipped with RGB cameras to capture high-resolution images of fruit trees from multiple angles. The images were processed using a novel deep convolutional neural network (DCNN) architecture called YOLOv5-v1, which was specifically modified to improve detection accuracy. The methodology involved the real-time synchronization of video streams from multiple UAVs, dynamic image capture, and the application of a filtering process to ensure precise object detection. The system was tested in various environmental conditions, and performance metrics such as precision, recall, and mean average precision (mAP) were used to evaluate its effectiveness.

The study's outcomes demonstrated that the proposed YOLOv5-v1 model achieved a high mean average precision of 86.8% in fruit detection and counting, surpassing existing technologies. The system maintained a low false positive rate of 14.7% and a false negative rate of 18.3%, even under challenging weather conditions.

These results highlight the system's potential for improving real-time fruit recognition and monitoring in orchard management, contributing significantly to precision agriculture.

Agricultural technology developers, orchard managers, and researchers in precision agriculture can apply these results to enhance automated fruit detection and yield estimation processes.

Technologies/methods used

• Multi-UAV imaging
• RGB cameras
• YOLOv5-v1 deep learning model
• Real-time video stream synchronization
• Kalman filter for object tracking

For further details, refer to the research article by Melnychenko et al. (2024) .

On the chart above: (a) the original YOLOv5 and (b) the proposed YOLOv5-v1. The left panel (a) shows the conventional structure of YOLOv5 with repeated bottleneck CSP layers, while the right panel (b) illustrates an augmented design of YOLOv5-v1 with additional bottleneck CSP layers and squeeze-and-excitation (SE) layers within the backbone and neck, ending in detection layers. This comparative layout underscores the modifications of the proposed YOLOv5-v1 aimed at improving the feature extraction and inference performance. Source: Melnychenko et al., 2024

On the scheme above - the figure presents two object detection processing pipelines and compares (a) a conventional approach of the original YOLOv5 with (b) an enhanced implementation of the proposed YOLOv5-v1. Both start with an input image sliced into multiple segments, proceed through the convolution and concatenation layers, and conclude with output images that have undergone batch normalization and activation functions. At the same time, the modified version introduces new dimension values for the input RGB channel (3 × 640 × 640) and feature maps (3 × 320 × 320) in the slice layer and suggests the Hardswish activation function in the focus module, hinting at efficiency gains. Source: Melnychenko et al., 2024

On the scheme above - (a) YOLOv5 with a shortcut connection and multiple convolutional layers with batch normalization (BN) and Hardswish activation functions, leading to a concatenated layer and output; (b) YOLOv5-v1 with one convolutional layer removed, leading to fewer parameters in the module. Source: Melnychenko et al., 2024

On the photos above: in panel (a), one apple at the bottom is circled in yellow, suggesting an omission in detection. In contrast, panel (b) shows the same scene without the omission, indicating a refined detection process. Source: Melnychenko et al., 2024

In the photos above: figure (a) exhibits its robust detection amidst cloud cover, (b) shows its accuracy with side light, (c) reveals the impact of backlighting on detection performance, and (d) indicates the challenges and potential overexposure when in direct sunlight. Green and blue boxes in these figures represent true positive and false positive cases, respectively. Source: Melnychenko et al., 2024

?? What's Next in Citrus Farming?

In our next edition of AI for citrus crops we will look at disease detection with UAVs.

?? How interesting is this topic for you?

?? Let us know in the comments below

?? Thank you for your time today

?? Stay in touch and... grow!

Wishes of good harvests of citrus fruits,

Maryna Kuzmenko, Ph.D ???? , Chief Citrus Eater at Petiole Pro

Photo credit for the cover of this edition:

Vijayakumar, V., Ampatzidis, Y., & Costa, L. (2023). Tree-level citrus yield prediction utilizing ground and aerial machine vision and machine learning. Smart Agricultural Technology, 3, 100077. https://doi.org/10.1016/j.atech.2022.100077 .

References for this edition of "AI for Citrus Orchards"