ASABE Technical Library - Abstract
Member and Access Notice
UAV-Based Yield Prediction and Phenology-Structured Feature Learning in Blackberry
Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org
Citation: 2026 ASABE Annual International Meeting 2601014.(doi:10.13031/aim.202601014)Authors: Akwasi Tagoe, Cengiz Koparan, PhD., Margaret Worthington, PhD., Donald Johnson, PhD., Dongyi Wang, PhD.
Keywords: Blackberry (Rubus spp.), Computer vision, High-throughput phenotyping, Precision horticulture; Statistical and machine learning models; UAV remote sensing; Yield prediction.
Abstract. Timely yield estimation in blackberry (Rubus spp.) is constrained by the labor-intensive nature of end-of-season harvest records, wide yield variation across genotypes, and the complexity of a multi-week bloom-to-harvest window. The objective of this study was to model blackberry yield using phenology-structured UAV-derived features within statistical and machine learning frameworks. UAV RGB imagery was evaluated as a pre-harvest phenotyping tool across two seasons: 2024 (primary; 11 weekly flights, N = 38 plots, 13 genotypes) and 2023 (supplementary; 4 post-peak flights, N = 44 plots) in a breeding trial containing floricane- and primocane-fruiting blackberry. Floral, canopy structural, and spectral features were extracted weekly using YOLOv10s-based flower detection, particle segmentation, and RGB vegetation indices (RGBVI, VARI, NGRDI, ExG), and summarized as peak, seasonal mean, and area-under-the-curve metrics. Multiple linear regression (MLR), polynomial regression, generalized additive models (GAM), and a linear mixed-effects model (LME) were fitted alongside machine learning models (Random Forest, XGBoost, SVM, KNN). MLR explained 68% of the variance in yield (R² = 0.68). XGBoost achieved the highest LOGO-CV performance (R² = 0.46; RMSE = 4,851 kg ha⁻¹), followed by Random Forest (R² = 0.323). Floral predictors dominated model performance, with vegetation features secondary and spectral indices minimal. Cross-year transfer (train 2024, test 2023) reduced accuracy (R² = 0.038), indicating limited generalizability. A phenology-informed framework for blackberry yield prediction was developed, with genotype-stratified evaluation of UAV-derived floral and canopy features. UAV-based phenotyping captures key components of yield formation, but robust prediction across seasons requires multi-season data and aligned feature extraction to yield.
(Download PDF) (Export to EndNotes)