# DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning

**Authors:** Alex Olsen, Dmitry A. Konovalov, Bronson Philippa, Peter Ridd, Jake C. Wood, Jamie Johns, Wesley Banks, Benjamin Girgenti, Owen Kenny, James Whinney, Brendan Calvert, Mostafa Rahimi Azghadi, Ronald D. White  
**arXiv:** [1810.05726](https://arxiv.org/abs/1810.05726)  
**Date:** 2018-10-09  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.05726v3/

## Key points

- Robotic weed control has potential to increase productivity in agriculture
- Research has focused on croplands, neglecting weed management challenges in rangeland stock farming
- DeepWeeds dataset created, first large-scale public image dataset of weed species from Australian rangelands
- Dataset consists of 17,509 labeled images of eight nationally significant weed species across eight locations in northern Australia
- Enables development of robust classification methods for robotic weed control
- Baseline classification performance achieved using Inception-v3 and ResNet-50 deep learning models: 95.1% and 95.7% average accuracies respectively
- Real-time performance demonstrated with ResNet-50 architecture, average inference time of 53.4 milliseconds per image
- Results pave the way for future implementation of robotic weed control methods in Australian rangelands
- Approach addresses various factors of variation encountered in rangeland environments (illumination changes, rotation and scale variations, occlusion, dynamic backgrounds, geographical and seasonal variations)
- Research contributes significantly to advancing robotic weed control technology by providing comprehensive dataset and demonstrating strong classification performance using deep learning models
- Technology has potential to enhance productivity and sustainability in agriculture by accurately identifying and managing weed species in rangeland environments

## Summary

Robotic weed control has gained attention in recent years for its potential to increase productivity in agriculture. Most research in this field has focused on developing robotics for croplands, overlooking the weed management challenges faced by rangeland stock farmers. To address this issue, the authors of this study have created the DeepWeeds dataset, which is the first large-scale public image dataset of weed species from the Australian rangelands. The dataset consists of 17,509 labeled images of eight nationally significant weed species found across eight locations in northern Australia. This dataset enables the development of robust classification methods that can make robotic weed control feasible. In this paper, the authors present a baseline for classification performance on the DeepWeeds dataset using two benchmark deep learning models: Inception-v3 and ResNet-50. These models achieved impressive average classification accuracies of 95.1% and 95.7%, respectively. The authors also demonstrate real-time performance of the ResNet-50 architecture, with an average inference time of 53.4 milliseconds per image. The results obtained from this study are highly encouraging and pave the way for future implementation of robotic weed control methods in Australian rangelands. The authors emphasize that their approach addresses several factors of variation commonly encountered in such environments, including illumination changes throughout the day, rotation and scale variations of target weeds during photography, occlusion caused by vegetation or other objects, dynamic backgrounds, and geographical and seasonal variations in plant life. Overall, this research contributes significantly to advancing robotic weed control technology by providing a comprehensive dataset and demonstrating strong classification performance using deep learning models. By accurately identifying and managing weed species in rangeland environments, this technology has the potential to greatly enhance productivity and sustainability in agriculture.