UAVs and Machine Learning Revolutionising Invasive Grass and Vegetation Surveys in Remote Arid Lands

Juan Sandino, Felipe Gonzalez, Kerrie L. Mengersen, Kevin J. Gaston

TL;DR

This paper presents a novel pipeline for mapping invasive grasses in remote arid lands using Unmanned Aerial Vehicles (UAVs) and machine learning. By integrating high-resolution RGB imagery with an eXtreme Gradient Boosting (XGBoost) classifier, the system achieves high accuracy in detecting and segmenting invasive species like buffel grass and spinifex, offering a cost-effective and efficient alternative to traditional survey methods.

What Problem Does This Paper Address?

Monitoring invasive grasses and vegetation in remote, often harsh, arid environments is a significant challenge. Traditional methods like ground surveys are costly, time-consuming, and can be dangerous. Satellite imagery, while useful, may lack the necessary ground sampling resolution, and cloud cover can hinder data acquisition. Manned aircraft surveys also face limitations. There's a clear need for accurate, efficient, and accessible surveillance methods to quantify invasion rates, track vegetation spread, and inform control strategies. This paper aims to fill this gap by developing an automated system for this purpose.

Methods

The core of this research is a four-component pipeline: Acquisition, Preprocessing, Training, and Prediction.

  • Acquisition: High-resolution RGB images were captured using a Canon EOS 5DsR camera mounted on a DJI S800 EVO UAV. The UAV was flown at a low altitude (approx. 67m) with significant overlap (80% forward, 50% side lap) to ensure detailed imagery with a Ground Sampling Distance (GSD) of about 1 cm/pixel.
  • Preprocessing: Raw images were orthorectified and processed using Agisoft PhotoScan Pro. Due to large file sizes, the orthomosaic was split into smaller tiles. Representative samples were then extracted and manually labelled using GIMP, categorizing pixels into six classes: buffel grass, soil/road, bushes, shadow, dry vegetation, and spinifex.
  • Training: A supervised machine learning approach was employed, specifically using the eXtreme Gradient Boosting (XGBoost) algorithm. The labelled data was converted into a feature array, incorporating color information (HSV color space) and texture features derived from 2D variance filters applied to hue, saturation, and grayscale images. Hyperparameters for the XGBoost classifier were tuned using a grid search with 10-fold cross-validation.
  • Prediction: The trained XGBoost model was then applied to the orthorectified imagery to predict the class of each pixel, generating a segmentation map of invasive grasses and other vegetation.

Datasets

The study utilized data collected from four flight campaigns in Cape Range National Park, Western Australia. Over 500 raw images were processed to create a large orthomosaic. From this, 342,626 pixel-wise samples were extracted and labelled into six classes for training and testing the machine learning model. The test dataset consisted of 85,657 labelled pixels for validation. No new benchmark datasets were introduced; rather, the paper focuses on applying existing techniques to a novel application and dataset.

Key Insights and Ideas

  • UAVs as a Game-Changer: The study highlights how UAVs, coupled with high-resolution RGB cameras, offer a practical and cost-effective solution for detailed vegetation mapping in remote areas, overcoming limitations of satellite and manned aircraft.
  • XGBoost's Robustness: The eXtreme Gradient Boosting algorithm proved highly effective, demonstrating robustness against variations in illumination, object rotation, occlusion, and background clutter.
  • Feature Importance: Hue, value, and saturation were found to be the most significant features for classification, accounting for up to 65% of the model's decision-making. Variance filters also played a crucial role, particularly for distinguishing textured objects like bushes and spinifex.
  • Feasibility of RGB: The research successfully demonstrates that high-resolution RGB imagery, without the need for multispectral or hyperspectral sensors, can be sufficient for accurate invasive grass detection when combined with advanced machine learning.

Key Results

The proposed pipeline achieved impressive results:

  • Individual Class Detection: 97% detection rate for buffel grass and 96% for spinifex.
  • Global Multiclass Detection: A remarkable 97% overall pixel-wise detection rate.
  • Classification Report: Precision rates ranged from 95.60% (Buffel) to 99.95% (Shadow), with recall rates from 84.15% (Bushes) to 99.99% (Shadow). The mean precision was 97.32% and the mean recall was 95.76%.
  • F-Score: The mean F-score across all classes was 96.54%, indicating a strong balance between precision and recall.
  • Robustness: The results were robust against illumination changes, object rotation, occlusion, background cluttering, and floral density variation.

Limitations and Future Work

  • "Bushes" Class Instability: The paper acknowledges that the "bushes" class exhibited instability, with misclassifications sometimes appearing as image noise. This is attributed to the visual similarity of bushes to other vegetation in RGB imagery and potential challenges in manual labelling.
  • Illumination Variation: While the current results are robust to some illumination changes, the study notes that acquired data is insufficient to fully test the approach under varying illumination conditions (e.g., different times of day, cloudy weather). Further research is needed to assess performance under these conditions.
  • Unsupervised Methods: Future work will focus on integrating unsupervised and supervised methods for vegetation data labelling to reduce processing times.
  • Real-time Applications: Enhancements are needed for real-time applications, including optimizing orthomosaic processing and improving software integration into a single, unified solution.
  • Algorithm Efficacy: Further analysis is recommended to compare the efficacy of supervised versus unsupervised algorithms for labeling vegetation and invasive grass species, and to integrate the best approaches into the pipeline.

Summary updated

Abstract

The monitoring of invasive grasses and vegetation in remote areas is challenging, costly, and on the ground sometimes dangerous. Satellite and manned aircraft surveys can assist but their use may be limited due to the ground sampling resolution or cloud cover. Straightforward and accurate surveillance methods are needed to quantify rates of grass invasion, offer appropriate vegetation tracking reports, and apply optimal control methods. This paper presents a pipeline process to detect and generate a pixel-wise segmentation of invasive grasses, using buffel grass (Cenchrus ciliaris) and spinifex (Triodia sp.) as examples. The process integrates unmanned aerial vehicles (UAVs) also commonly known as drones, high-resolution red, green, blue colour model (RGB) cameras, and a data processing approach based on machine learning algorithms. The methods are illustrated with data acquired in Cape Range National Park, Western Australia (WA), Australia, orthorectified in Agisoft Photoscan Pro, and processed in Python programming language, scikit-learn, and eXtreme Gradient Boosting (XGBoost) libraries. In total, 342,626 samples were extracted from the obtained data set and labelled into six classes. Segmentation results provided an individual detection rate of 97% for buffel grass and 96% for spinifex, with a global multiclass pixel-wise detection rate of 97%. Obtained results were robust against illumination changes, object rotation, occlusion, background cluttering, and floral density variation.
BibTeX
@inproceedings{sandino2018uavs,
  author = {Juan Sandino and Felipe Gonzalez and Kerrie L. Mengersen and Kevin J. Gaston},
  title = {{UAVs and Machine Learning Revolutionising Invasive Grass and Vegetation Surveys in Remote Arid Lands}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=-XsO5aywEq},
}

Loading related content...

Figure