CottonWeeds: Empowering precision weed management through deep learning and comprehensive dataset
In sustainable crop production, effective weed management is crucial for agricultural productivity. Although deep learning techniques for weed detection have gained attention, their adoption is hindered by the lack of comprehensive datasets. Addressing this gap, we present ‘CottonWeeds,’ a meticulou...
Gespeichert in:
Veröffentlicht in: | Crop protection 2024-07, Vol.181, p.106675, Article 106675 |
---|---|
Hauptverfasser: | , |
Format: | Artikel |
Sprache: | eng |
Schlagworte: | |
Online-Zugang: | Volltext |
Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Zusammenfassung: | In sustainable crop production, effective weed management is crucial for agricultural productivity. Although deep learning techniques for weed detection have gained attention, their adoption is hindered by the lack of comprehensive datasets. Addressing this gap, we present ‘CottonWeeds,’ a meticulously curated database with 7578 images of horse purslane (Trianthema portulacastrum L.) and purple nutsedge (Cyperus rotundus L.), two common weed species in cotton (Gossypium hirsutum L.) fields. Captured under diverse conditions in the Indian subcontinent during the 2021 growth season, these labeled images constitute a valuable resource. We rigorously evaluated 11 state-of-the-art image classification models using deep transfer learning (DTL), achieving exceptional accuracy exceeding 93% within a 3-h training period. MobileNet emerged as the top performer, reaching 95.43% accuracy. Challenges arose with limited purple nutsedge training samples, impacting conventional models. To address this, we introduced a weighted cross-entropy loss function, significantly enhancing purple nutsedge classification. For instance, VGG16's purple nutsedge accuracy improved from 92.00% to 96.57%. Additionally, we assessed the dataset's utility for object detection, employing pre-trained YOLOv5 models. YOLOv5x exhibited exceptional performance, with a mAP_0.5:0.95 score of 72.5% and a mAP_0.5 score of 87.4%, proving invaluable for weed identification in cotton fields. Our study underscores CottonWeeds' potential as a foundational training platform for real-time, in-field weed recognition models.
[Display omitted]
•Developed benchmark with 11 deep learning models for weed classification.•Curated dataset: 7578 images from Indian Subcontinent cotton fields.•Demonstrated efficacy of weighted loss function for improved accuracies.•Explored real-time weed detection and removal using YOLOv5 models (v5x, v5m, v5l). |
---|---|
ISSN: | 0261-2194 1873-6904 |
DOI: | 10.1016/j.cropro.2024.106675 |