Rice tiller number estimation based on an improved Swin-UNet model and multi-feature fusion

TL;DR

This paper introduces a new method for estimating rice tiller numbers using an improved Swin-UNet model and multi-feature fusion, enhancing accuracy for high-yield rice breeding.

Contribution

The novel contribution is an improved Swin-UNet model combined with PSO-optimized XGBoost for accurate rice tiller number estimation.

Findings

The improved Swin-UNet model achieved 92.5% segmentation accuracy, 7.2% higher than U-Net.

The PSO-XGBoost model using 12 features achieved R²=0.85 and RMSE=0.35 for tiller number estimation.

Application on 576 breeding plots produced accurate tiller number maps for smart breeding.

Abstract

Rice early tillering characteristics are key indicators for high-yield breeding, with tiller number and tillering rate as core parameters. High-throughput, temporal, and precise monitoring of tiller numbers via drone digital imagery provides quantitative support for tillering trait screening in breeding, serving as an important auxiliary tool for smart breeding. However, during the early tillering stage, complex backgrounds (e.g., water bodies, soil) and small, dense breeding plots pose challenges to high-throughput rice plant extraction and accurate tiller number estimation. To address this, this study proposes a rice tiller number estimation method based on an improved Swin-UNet model and multi-feature fusion. A PSO-optimized XGBoost model was constructed for tiller number estimation by integrating selected features. Experimental results show that the improved Swin-UNet model achieved a segmentation accuracy of 92.5% (7.2% higher than U-Net), and the PSO-XGBoost model, using 12 features (10 morphological and 2 color), yielded R²=0.85 and RMSE = 0.35. Application verification on 576 untrained breeding plots generated tiller number thematic maps, providing data support for germplasm tillering trait identification and advancing smart breeding.

Figures