From segmentation to classification: Morphological phenotype extraction and classification analysis of tiny poplar seeds using the MP-Seed segmentation algorithm

Abstract

Extracting poplar seed morphological phenotypes is a core task in modern poplar breeding research. Accurate seed image segmentation is crucial for phenotype extraction and data quality. However, the small size of poplar seeds and their tendency to form dense clusters challenge the accuracy of current segmentation methods. Unlike current approaches that struggle with small-target segmentation and boundary delineation, this study develops the MP-Seed semantic segmentation algorithm, which combines a small-target attention module (based on Layer Across Feature Map Attention) with a multi-task learning mechanism that integrates boundary features. This novel integration targets small-seed key regions, fuses boundary features, and refines predictions to precisely segment densely clustered seeds, achieving superior accuracy and fine-grained delineation compared to current single-task methods. To address low efficiency and accuracy in poplar seed morphological phenotype extraction, this study further proposes a high-throughput extraction method leveraging the MP-Seed algorithm. To analyse the phenotypic data, an SVM classification model classifies eight types of poplar seeds. Experimental validation shows that the MP-Seed algorithm outperforms current methods on the test set, achieving Seed_IoU of 94.1 %, mIoU of 97.2 %, and Reference_IoU of 97.6 %. The high-throughput phenotyping method measures seed length and width with relative errors within 2.72 % versus manual measurements and extracts ten morphological traits at about 18.3 seeds per second. The overall classification accuracy reaches 91.1 %. Overall, this study provides technical support for accurate poplar seed segmentation and efficient morphological phenotype extraction, offering a valuable reference for other seed morphological phenotype research and analysis.

Plant phenotyping relevance

ポプラ種子画像のセグメンテーションアルゴリズムと高スループット形態形質抽出法を開発・検証しており、植物フェノタイピング手法が研究の中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.