Location: Genetics and Sustainable Agriculture Research
Project Number: 6064-21000-017-001-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Aug 1, 2026
End Date: Jul 31, 2028
Objective:
The primary objective of this project is to develop a computationally efficient, anchor-free deep learning model for accurate cotton bloom detection and counting from high-resolution UAV imagery. The proposed framework will be specifically designed to address the challenges of small-object detection, extreme class imbalance, and complex field environments. A secondary objective is to generate spatially explicit bloom maps to quantify flowering patterns and their relationships with plot-level yield performance.
Approach:
High-resolution RGB imagery was collected over experimental cotton plots. These images highlight key technical challenges of bloom detection, including small object size, background dominance, and canopy occlusion. A point-level annotation strategy will be employed to label bloom centers, significantly reducing labeling effort compared to bounding-box annotation. Semi-automated model-assisted labeling will further accelerate dataset construction. Data augmentation techniques, including illumination variation, rotation, and canopy occlusion simulation, will be applied to improve model robustness. In addition, end-of-season yield data were collected for each plot, providing ground-truth measurements for subsequent bloom–yield analysis. Using these data, we will develop a novel anchor-free cotton bloom detection framework based on a Vision Transformer (ViT) backbone, specifically redesigned for small-object localization in high-resolution UAV imagery. Although ViTs provide strong global representation, direct application is inefficient for background-dominated agricultural images where cotton blooms are extremely sparse. To address this, Transformer multilayer perceptrons will be replaced with Kolmogorov–Arnold Network (KAN)-based modules, and a Padé KAN (PKAN) design will improve nonlinear modeling capacity and numerical stability. We will redesign the attention mechanism using additive operations to reduce computational complexity from O(N²) to O(N). Spatial and channel attention will enhance bloom features and suppress background noise. The model follows an encoder–decoder architecture, where a CNN backbone extracts multi-scale features that are embedded and passed to the Transformer encoder, and a decoder with learnable instance queries directly predicts bloom center locations. All collected imagery, annotated datasets, trained models, and processing code will be managed following FAIR (Findable, Accessible, Interoperable, Reusable) data principles. Raw UAV imagery and derived products will be archived in publicly accessible repositories with comprehensive metadata, enabling reproducibility and facilitating adoption by the broader research community. Throughout our work, SCINet computing resources will be used for performing deep learning computations and data storage. The PI, Dr. Jixiang Wu at USDA ARS, intends to integrate remote sensing technology into cotton field evaluations to improve ground truth trait prediction. While his expertise is in genetic and breeding design, field experimentation, and quantitative genetics, remote sensing data processing lies outside his formal training. To address this critical gap, our co PI,at the University of Wisconsin, provides essential expertise and extensive experience in precision agriculture, including satellite based remote sensing, drone imaging platform development, multi source data fusion, and the application of artificial intelligence and machine learning to agricultural decision support. Their contributions are vital for delivering the remote sensing components of this project.