Taxonomical loss for weed seedlings image classification

TL;DR

This paper introduces a new deep learning method to better classify weed seedlings in early growth stages, which could help reduce pesticide use in agriculture.

Contribution

A novel taxonomic loss function for few-shot learning is proposed, which improves clustering and classification of weed seedlings.

Findings

The taxonomic loss function improved clustering performance based on Silhouette scores compared to triplet loss.

The method enhanced the identification of weed seedlings at early growth stages.

The taxonomic loss did not consistently improve all deep learning model architectures.

Abstract

Accurate classification of weed seedlings is a key challenge to address in order to advance precision weed management and reduce pesticide use. In this study, an original image dataset, the Weed Phenological Dataset (WPD), with annotated early growth stages is introduced. Furthermore, a novel deep learning taxonomic loss function for few-shot learning was evaluated. Using hierarchical structures to direct the classification process, this taxonomic loss function introduces dynamic margins during the computation. In experiments using the taxonomic loss using the ResNet-50 architectures across different plant image datasets, this taxonomic approach allowed better clustering according to the Silhouette scores when compared with triplet loss using 100 images per class. It also led to better identification of weed seedlings at early growth stages. Although the new taxonomic loss did not consistency improve classification results for all the deep learning model architectures, it open new research avenues for the robotization of agriculture.

The online version contains supplementary material available at 10.1038/s41598-025-33961-0.

Figures