Written by Dattatray Bhalekar, Juan Munguia de la Cruz, Bernardita Sallato, Lav Khot, Washington State University, and Michael Meyer, Ines Hanrahan, Tory Schmidt, Washington Tree Fruit Research Commission, May 20, 2025
Crop load monitoring
Crop load management is a fundamental practice to ensure fruit quality and consistent yields in apple production. One of the most critical steps of crop load management is the assessment of bloom and green fruit density to guide thinning strategies. Thus, being able to precisely monitor crop load can aid growers making better farming decisions. Precise crop load data and management is also useful to prevent biennial bearing, estimate nutrient needs, and determine harvest strategies with the packing house and sales desk. With conventional labor-intensive hand counting methods, only a subset of trees can be sampled to estimate crop load and decide management for an entire block, which can impact orchard variability and reduce efficiencies. Thus, there is a need for more efficient and precise techniques for crop load monitoring. To address these issues, several commercial aerial or ground-based machine vision solutions have been on the horizon and are being validated in the WSU Smart Orchard testbed near Mattawa, WA. This article focuses on one of the ground-based crop load mapping platforms, i.e., Vivid Machines Inc.
This platform is being validated in two WA orchards. The Vivid XV3 Vision System uses a multispectral sensor to capture multiple bands in the visible and near-infrared (VIS-NIR) spectral range, a GPS receiver, and an integrated proprietary data processing algorithm. The system can be universally mounted on any farm vehicle to scan the block at 5-10 mph (Figure 1). Vivid Machines claims that the onboard vision system can scan, process, and provide real-time estimates for up to 20,000 trees/hour. They can generate variability maps based on estimated crop parameters at tree-level resolution. The variability maps (Figure 2), processed data, and scanning reports can be accessed remotely using a smartphone with the Vivid Control app or cloud-based dashboards.
Methods
In 2024 season, two independent trials were conducted in WA commercial apple orchards. The first on a 4-year-old ‘Cosmic Crisp® (WA 38)’ orchard near Mattawa, and the second on a 20-year-old organic ‘Honeycrisp’ orchard near Naches. Both are trained to a tall spindle architecture in a vertical trellis system.
To validate the estimates provided by Vivid Machines, crop parameters including blossom cluster counts, fruit count, and fruit size (diameter, mm) were manually measured at key growth stages throughout the season. At the Mattawa site, ground truth blossom cluster count was collected at three zones within the orchard with 12-15 trees per zone. Additionally, fruit diameter and fruit count were collected at four different zones with five trees per zone. Five fruits were randomly selected from each tree to measure fruit diameter using a vernier caliper device. While at the Naches site, three zones with two trees per zone were selected to collect blossom clusters, fruit count, and fruit diameter.
The absolute percentage error (APE, %) was calculated using Eq.1, based on estimated average blossom clusters/ fruit count per tree for the entire block and actual blossom clusters/ fruit counts from ground truth trees.
The same equation was used to quantify APE in fruit sizing based on average estimated fruit diameter (mm) per tree for the entire block and ground truth trees.
Results
Mattawa (Cosmic Crisp®): An average blossom cluster count per tree (mean ± Standard Deviation) estimated by Vivid Machines was 108.6 ± 35.6 clusters, compared to actual counts of 131.4 ± 56.8 clusters, indicating 17.3 % APE.
Vivid Machines scans indicated an APE of 15.5% over the entire season, with error decreasing as the season progressed (5.3% APE at peak performance). Similarly, for estimating fruit size (diameter > 20 mm), the overall average APE was 3.8%, with the lowest error of 2.3% observed on July 18, 2024 (Figure 3).
Naches (Honeycrisp): Vivid Machines estimated 176 ± 52.2 blossom clusters per tree, corresponding to an APE of 42.5% when compared to the actual count of 306 ± 73.8 clusters per tree. Early in the season (May 14), when fruit diameters were less than 10 mm, fruit count estimation showed a high APE of 51.9%. However, as fruit diameter increased to over 20 mm, the error reduced to 38.4% on June 15 and reached a minimum of 28.5% at harvest. For fruit sizing, scans between May 8 and 14 exhibited an average APE of 23%, which significantly decreased to 0.5% by June 15 and remained low at 3% for the pre-harvest scan (Figure 4).
Key findings
- Overall, Vivid Machines demonstrated a lower error in estimating fruit size, compared to blossom cluster or fruit count.
- Reduced error in fruit size estimation was observed when the fruit diameter was larger than 20 mm in mid and late season.
- The estimation error for blossom clusters and fruit count was relatively higher at the Naches site compared to the Mattawa site, possibly due to variability between the blocks.
Accurately capturing spatial variability in various crop parameters using digital technologies similar to Vivid Machines can be a valuable grower decision support tool. For example, blossom or fruitlet density variability can be used to create prescription maps for precision chemical thinning using variable rate sprayers. Crop load variability data can also be used to decide harvest time and bin placement within the orchard.
Ongoing efforts
Since the 2024 season, thanks to grower feedback and research projects such as this one, several enhancements have been made to Vivid Machines’ digital vision system to improve its performance and accuracy of count predictions, particularly for trees with very high or very low counts. Models are now trained on a larger, more diverse dataset with improved training techniques, resulting in better generalization. A key upgrade has been the transition to processing high-resolution images in real-time, replacing the previous reliance on heavily down-sampled imagery change which may improve prediction quality.
In 2025 season, the WSU team will focus on validating similar commercial precision crop load monitoring technologies at Smart Orchard Testbeds (Mattawa and Zillah). Efforts are being made to develop prescription maps that can be ingested by the variable rate sprayers for precision blossom thinning applications. One such commercial collaboration being explored is with Aurea Imaging Pvt. Co., doing scans using their TreeScout platform and generating maps to feed into a Turbo-mist sprayer (from Slimline Mfg.) that is being retrofitted with a Raven Precision Spray System. This integration is being piloted by Innov8.Ag Pvt Co., and Burrows Tractor Inc., with the WSU team in a support role.
Contacts
Bernardita Sallato
Associate Professor
WSU Tree Fruit Extension
b.sallato@wsu.edu
(509) 439-8542
Lav Khot
Associate Professor of Precision Agriculture
WSU CPAAS and AgWeatherNet
lav.khot@wsu.edu
(509) 786-9302
Ines Hanrahan
Washington State Tree Fruit Research Commission
hanrahan@treefruitresearch.com
(509) 669-0267
Funding and acknowledgements
These efforts are funded in part by the USDA-NIFA Specialty Crop Research Initiative (SCRI) project (Precision Apple Cropload Management, PACMan) and the Washington Tree Fruit Research Commission (WTFRC). Authors would like to extend sincere appreciation to Vivid Machines Inc. for their collaboration and in-kind mapping campaigns at Smart Orchard Testbeds. We would also like to thank grower cooperators, Northwest Farm Management LLC., and Allan Bros, Inc.
Additional Reading
Bhalekar, D., Munguia, J., Khot, L., & Sallato, B. (2024). Precision Crop Monitoring Techniques at WSU Smart Apple Orchard Testbed. WSU Smart Apple Orchard, Washington State University. https://smartorchard.wsu.edu/precision-crop-monitoring-techniques-at-wsu-smart-apple-orchard-testbed/
https://smartorchard.wsu.edu/Wallis, A., Clements, J., Sazo, M. M., Kahlke, C., Lewis, K., Kon, T., … & Robinson, T. (2023). Digital Technologies for Precision Apple Crop Load Management (PACMAN) Part I: Experiences with Tools for Predicting Fruit Set Based on the Fruit Growth Rate Model. NY Fruit Q, 31, 8-13.
Fruit Matters articles may only be republished with prior author permission © Washington State University. Reprint articles with permission must include: Originally published by Washington State Tree Fruit Extension Fruit Matters at treefruit.wsu.edu and a link to the original article.
