Source-linked AI summary
DeepWeeds: A Multiclass Weed Species Image Dataset for Deep Learning
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
TL;DR
Robust weed-species classification in complex Australian rangeland environments remains underrepresented in robotic weed-control research. This paper introduces the DeepWeeds dataset and benchmarks deep-learning classifiers, achieving 95.7% average accuracy with ResNet-50 and 53.4 ms inference using TensorRT.
Problem
Weed-species classification in Australian rangelands is underrepresented, despite challenging terrain, backgrounds, lighting, and occlusion that complicate robotic weed control.
Method
The study collects the in-situ DeepWeeds dataset and benchmarks modified Inception-v3 and ResNet-50 classifiers for weed-species recognition and real-time inference.
Results
95.7% average top-1 classification accuracy was achieved by ResNet-50, with TensorRT inference averaging 53.4 ms per image.
Takeaways & Limitations
The results provide a strong baseline for rangeland weed classification and meet the stated real-time inference target for the robotic weed-control application.
Takeaways & Limitations
False positives can misclassify native plants as weeds, potentially harming native ecosystems and wasting herbicide.
Abstract
from arXiv · showhide
Robotic weed control has seen increased research of late with its potential for boosting productivity in agriculture. Majority of works focus on developing robotics for croplands, ignoring the weed management problems facing rangeland stock farmers. Perhaps the greatest obstacle to widespread uptake of robotic weed control is the robust classification of weed species in their natural environment. The unparalleled successes of deep learning make it an ideal candidate for recognising various weed species in the complex rangeland environment. This work contributes the first large, public, multiclass image dataset of weed species from the Australian rangelands; allowing for the development of robust classification methods to make robotic weed control viable. The DeepWeeds dataset consists of 17,509 labelled images of eight nationally significant weed species native to eight locations across northern Australia. This paper presents a baseline for classification performance on the dataset using the benchmark deep learning models, Inception-v3 and ResNet-50. These models achieved an average classification accuracy of 95.1% and 95.7%, respectively. We also demonstrate real time performance of the ResNet-50 architecture, with an average inference time of 53.4 ms per image. These strong results bode well for future field implementation of robotic weed control methods in the Australian rangelands.
Introduction
Robotic weed control could improve agricultural productivity by reducing labour and herbicide use, but robust weed detection and classification remain obstacles, especially in rangeland environments. DeepWeeds addresses this gap with a human-labelled, in situ dataset of eight weed species collected across northern Australia.
- Motivation: Robotic weed control may reduce labour costs and herbicide usage through efficient selective application to weed targets.Improved weed-control efficacy could have substantial economic impact.
- Challenge: Detection and classification remain significant obstacles to commercial development and industry acceptance of robotic weed-control technology.Existing approaches use image-based, spectrum-based, and spectral methods with varied success.
- Approach: Deep learning models, including Convolutional Neural Networks, are among the most promising methods for automatic plant recognition from leaf images.CNNs now dominate many computer-vision-related fields.
- Dataset gap: Most weed-classification datasets capture plants under perfect laboratory conditions, limiting their representation of natural environments.Such conditions support strong theoretical classification performance but differ from field settings.
- Rangeland context: Rangeland weed classification has been largely ignored despite remote, extensive, and challenging environments that differ from cropping applications.Cropping environments are often flat, homogeneous, and subject to controlled light conditions.
- Contribution: 17,509 labelled images cover eight weed species collected in situ from eight rangeland environments across northern Australia.The study also trains a deep-learning classifier and validates its real-time performance.
Methods
The methods were designed to collect a large, varied weed-image dataset under conditions resembling robotic field operation. They also established CNN baselines by adapting pretrained Inception-v3 and ResNet-50 models for nine-class weed classification.
- Dataset design: The dataset-collection design considered optical systems, scene variability, dataset size, weed targets and locations, negative samples, and image acquisition.These factors were selected to support robust classification for robotic weed control.
- Image acquisition: 10 km/hr vehicle speed and a 450×280 mm field of view allowed approximately 100 ms per image, requiring fast shutter speeds to limit motion blur.The system was intended to detect a weed before the next image became available for processing.
- Image acquisition: The WeedLogger used the target platform’s optical system, consistent acquisition, and GPS logging, combining a Raspberry Pi, high-resolution camera, machine-vision lens, and GPS receiver.The selected camera was a FLIR Blackfly 23S6C Gigabit Ethernet high-resolution colour camera with 1920 x 1200 px resolution.
- Dataset design: 1,000 images per target species and a 50:50 positive-to-negative image split at each location were established as dataset goals.The image-volume goal supports training high-complexity CNNs, while the class balance reduces overfitting to scene-level features.
- CNN baselines: Inception-v3 and ResNet-50 pretrained on ImageNet were modified by replacing their 1,000-neuron output layer with a 9-neuron fully connected layer.The modified networks were implemented using Keras with TensorFlow as the machine-learning framework.
- CNN baselines: A sigmoid neuron was used for each weed-specific output, and an image was assigned the class with the greatest sigmoid-activated probability.This output design accommodates images containing multiple weed species while selecting the most likely target class.
Results
The DeepWeeds dataset comprises 17,509 images spanning eight weed species and negative plant-life classes collected across northern Australia. ResNet-50 and Inception-v3 established strong classification baselines, while TensorRT enabled ResNet-50 to meet the real-time inference target.
- Dataset: 17,509 images comprise DeepWeeds, covering eight weed species and off-target or negative plant life collected across northern Australia.Images were collected from June 2017 to March 2018 using the WeedLogger in-field instrument.
- Dataset: Over 1,000 images were collected for each weed species, totaling over 8,000 positive-class images with evenly split positive and negative samples at each location.Negative samples contained neighbouring flora and backgrounds without the weed species of interest.
- Classification performance: 95.7% average top-1 classification accuracy was achieved by ResNet-50, compared with 95.1% for Inception-v3 across five cross-validated folds.ResNet-50 also outperformed Inception-v3 in precision and false positive rate.
- Classification performance: 97.6% average top-1 accuracy was ResNet-50’s strongest species-level result, achieved on the negative class; parkinsonia reached 97.2%.Other strong results included siam weed at 96.5%, parthenium at 95.8%, prickly acacia at 95.5%, and lantana at 95.0%.
- Real-time performance: 53.4 ms average pre-processing and inference time enabled ResNet-50 to meet the 100 ms-per-image real-time target with TensorRT, unlike standard TensorFlow.The target corresponds to processing at least 10 frames per second for the selective spot sprayer.
Conclusions
The work introduces DeepWeeds, a large in situ multiclass dataset of eight nationally significant Australian weed species collected across eight northern Australian locations. It establishes strong baseline classification performance and identifies directions for improving robustness, field deployment, and imaging-based classification.
- Dataset contribution: DeepWeeds is the first large multiclass weed-species image dataset collected entirely in situ from Australian rangelands.It contains eight weed species of national significance to Australia and spans eight geographic locations across northern Australia.
- Baseline performance: 95.1% average classification performance was achieved by Inception-v3 on the DeepWeeds dataset.This result is reported as a baseline using a convolutional neural network model.
- Baseline performance: 95.7% average classification performance was achieved by ResNet-50 on the DeepWeeds dataset.This result is reported alongside the Inception-v3 baseline.
- Future work: Future work includes improving classification accuracy and robustness under realistic conditions.The authors also anticipate further research into classifying rangeland weeds using the DeepWeeds dataset.
- Future work: Future work includes field implementation of the learning models in a prototype weed-control robot and investigating NIR spectroscopy and hyperspectral imaging.These directions target weed detection and species classification.