Accurate Crop Yield Estimation of Blueberries using Deep Learning and Smart Drones

Hieu D. Nguyen, Brandon McHenry, Thanh Nguyen, Harper Zappone, Anthony Thompson, Chau Tran, Anthony Segrest, Luke Tonon · 2025

Read Accurate Crop Yield Estimation of Blueberries using Deep Learning and Smart Drones by Hieu D. Nguyen, Brandon McHenry, Thanh Nguyen, Harper Zappone, Anthony Thompson, Chau Tran, Anthony Segrest, Luke Tonon (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-02344.txt.

Introduction

Nguyen and co-authors present a 2025 paper on counting blueberries from smart drones to forecast harvest size. Hand counts are slow and rough, so growers want camera-based estimates. Their pipeline pairs two detection models built on the YOLO family. One spots whole bushes from low flights, and the other counts single berries on a bush.

Smart flight control brings the drone in for safe, close side views of each bush. Cropping around the front bush lifted both precision and recall in their tests. The paper also maps field sampling plans for full surveys. Labeling such tiny fruit proved a real chore, and judging true counts stays tricky, they note.

Catalog record

CREATORHieu D. Nguyen, Brandon McHenry, Thanh Nguyen, Harper Zappone, Anthony Thompson, Chau Tran, Anthony Segrest, Luke Tonon
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-02344

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Accurate Crop Yield Estimation of Blueberries using Deep Learning and Smart Drones Hieu D. Nguyen, Brandon McHenry, Thanh Nguyen, Harper Zappone, Anthony Thompson, Chau Tran, Anthony Segrest, Luke Tonon Abstract We present an AI pipeline that involves using smart drones equipped with computer vision to obtain a more accurate fruit count and yield estimation of the number of blueberries in a field. The core components are two object-detection models based on the YOLO deep learning architecture: a Bush Model that is able to detect blueberry bushes from images captured at low altitudes and at different angles, and a Berry Model that can detect individual berries that are visible on a bush. Together, both models allow for more accurate crop yield estimation by allowing intelligent control of the drone’s position and camera to safely capture side-view images of bushes up close. In addition to providing experimental results for our models, which show good accuracy in terms of precision and recall when captured images are cropped around the foreground center bush, we also describe how to deploy our models to map out blueberry fields using different sampling strategies, and discuss the challenges of annotating very small objects (blueberries) and difficulties in evaluating the effectiveness of our models. 1 Introduction Precision agriculture using AI and autonomous drones has been shown to be an effective approach in not only estimating yield for many different crops but also in detecting and managing weeds and diseases. One popular specialty crop where accurate estimation of crop yield is important is the blueberry, where prediction early in the growing season is important in helping farmers make pricing decisions, hire a sufficient number of pickers, and inform their distributors of available supply before harvest time. One AI-based approach to estimating crop yield involves using deep learning models trained on images to detect fruits and vegetables; many such models have been developed for many of them, including blueberries, grapes, apples, and tomatoes. However, early models were trained on either simulated data [4] or close-up images of clusters of fruit (from hand-held cameras or mobile devices) and not the entire bush or vine [5-9]; thus, these models are inappropriate for industrial use where crop count and disease detection must be performed over a large field many acres in size. For example, a one-acre blueberry field can contain up to 1000 bushes, which would make walking through such a field to capture a good sample of close-up images quite time-consuming; thus, a “boots-on-the-ground” approach is impractical. On the other hand, using unmanned aerial vehicles (UAVs), in our case drones, to capture images of blueberry bushes from a farther distance presents a much more efficient approach to mapping large fields 1 1 1 We note that ground autonomous vehicles (or robots) offer an alternative solution; however, we shall not discuss them and their trade-offs in this paper since this topic has been well addressed in the literature and doing so will set us too far adrift from the focus of our work. . In particular, we consider smart drones programmed with computer vision, i.e., drones that have an onboard mini-computer to process images captured by an onboard camera, to identify a blueberry bush and position itself to capture an optimal view of the bush that maximizes the number of visible berries on the bush. In addition, we make a distinction between a smart drone vs an autonomous drone, where the former is capable of making its own decisions during flight, e.g., by altering its path to avoid collision, as opposed to an autonomous drone where its actions, including its flight path, is pre-programmed before take-off. For example, commercial drones that are equipped with collision-avoidance systems (based on technologies such as infrared, stereo vision, and LiDAR) and provide object-tracking capabilities would fall under our classification of a smart drone. In this work, we present a pipeline of object-detection models based on the YOLO (You Only Look Once [ 1 ] ) deep learning architecture to estimate the crop yield of a blueberry field by using smart drones programmed with these models to accurately capture images of blueberry bushes and detect the number of harvestable berries on each bush (Figure 1 shows an example of a blueberry bush). Our work is novel in its approach of detecting not only individual berries that are visible on an entire blueberry bush but also in detecting the bush itself in order to guarantee accurate drone position and image capture of the bush. We distinguish such a smart mission from an autonomous mission as follows: in the latter, a drone is programmed before the start of the mission to fly to predetermined points of a blueberry field, say, along a row of bushes but keeping itself some fixed distance away from the row (and high enough to clear neighboring rows); moreover, its onboard camera can be programmed to capture images of these bushes, but from a fixed angle of view. However, this does not always guarantee that the bushes will be fully captured by the drone’s camera. For example, this situation may occur if the drone’s position is blown off course due to wind or temporary loss of GPS. Thus, it is not possible for the drone to adjust its distance from the bushes or its camera’s angle of view in real time to compensate for this. In contrast, drones programmed with our Bush Model can fly a more intelligent mission where it is able to adjust its position and fly as close as possible alongside a bush (while maintaining a safe distance) in order to capture an optimal view of the bush so that berries on it appear as largest as possible (for more accurate detection). This also allows for the implementation of various sampling strategies, e.g., random stratified sampling of bushes, without having to know the exact GPS location of each bush, something that would be required for autonomous missions. On the other hand, coupled with real-time kinematic (RTK) positioning, our pipeline makes precision mapping of blueberry fields possible where the location of each bush can be geotagged using our Bush Model. Berry count can then calculated for each bush post flight using our Berry Model. Then given appropriate field data, which we describe in Section III, crop yield can be estimated more accurately than from current methods. The rest of the paper is divided as follows. In Section II, we discuss related works. In Section III, we present our proposed pipeline for mapping a blueberry field to obtain a more precise estimation of crop yield. In Section IV, we present experimental results for our models, which we hope provides baseline results for future researchers to compare their work against, and discuss challenges we faced in annotating tiny objects (blueberries) and how this impacted the effectiveness of our models. 2 Related Works Recent advances in computer vision, in particular object-detection models based on deep learning such as the YOLO architecture, have led to a proliferation of works in precision agriculture. More specifically, many of these works present highly accurate models to perform fruit detection and yield estimation. Since many articles have recently been published in this field (a 2020 survey article [ 2 ] reviewed 30 articles

Read the complete text