TY - JOUR
T1 - Bitou bush detection and mapping using UAV-based multispectral and hyperspectral imagery and artificial intelligence
AU - Amarasingam, Narmilan
AU - Kelly, Jane E
AU - Sandino , Juan
AU - Hamilton, Mark
AU - Gonzalez, Felipe
AU - Dehaan, Remy L
AU - Zheng, Lihong
AU - Cherry, Hillary
PY - 2024/4
Y1 - 2024/4
N2 - The use of Unmanned Aerial Vehicles (UAVs) for remote sensing (RS) of vegetation presents a valuable platform for weed monitoring, owing to the high spatial resolution of collected images. Accurate segmentation and mapping of weed spatial distribution plays a pivotal role in achieving effective management and ensures efficient and sustainable utilization of weed control measures. Furthermore, UAV-based RS provides a rapid way of assessing phenological development stages of weed species such as flowering and fruiting. These are often critical stages required for the separation of weed species from surrounding vegetation and are difficult to capture with traditional low resolution airborne and satellite RS imagery. Bitou bush is a shrub and a serious environmental weed of coastal areas of New South Wales (NSW). The primary objective of this study is to develop a model for bitou bush mapping from collected multispectral (MS) and hyperspectral (HS) imagery on location in NSW, Australia by employing various classical machine learning (ML) and deep learning (DL) techniques. The performance of Random forests (RF), Support vector machine (SVM), Extreme gradient boosting (XGB), and K-nearest neighbors (KNN) models is evaluated, achieving overall validation accuracies of 78%, 74%, 80%, and 69%, respectively for bitou bush detection using MS imagery. Subsequently, these models are assessed on HS data, resulting in overall validation accuracies of 77%, 86%, 86%, and 80% for RF, SVM, XGB, and KNN, respectively. Moreover, the DL U-Net model achieved an overall validation accuracy of 92%, outperforming the classical ML models in MS data segmentation tasks. The results of this study highlight the superior performance of the U-Net model in comparison to classical ML models in RS data segmentation, indicating the value of DL techniques for more accurate and robust RS applications such as bitou bush detection and mapping. The insights gained from this research will aid researchers and land managers select appropriate models based on the complexity and characteristics of their RS datasets. Moreover, the integration of UAV RS and artificial intelligence (AI) provide a valuable and efficient platform for bitou bush monitoring and management practices, ultimately enhancing the efficiency and sustainability of weed control efforts.
AB - The use of Unmanned Aerial Vehicles (UAVs) for remote sensing (RS) of vegetation presents a valuable platform for weed monitoring, owing to the high spatial resolution of collected images. Accurate segmentation and mapping of weed spatial distribution plays a pivotal role in achieving effective management and ensures efficient and sustainable utilization of weed control measures. Furthermore, UAV-based RS provides a rapid way of assessing phenological development stages of weed species such as flowering and fruiting. These are often critical stages required for the separation of weed species from surrounding vegetation and are difficult to capture with traditional low resolution airborne and satellite RS imagery. Bitou bush is a shrub and a serious environmental weed of coastal areas of New South Wales (NSW). The primary objective of this study is to develop a model for bitou bush mapping from collected multispectral (MS) and hyperspectral (HS) imagery on location in NSW, Australia by employing various classical machine learning (ML) and deep learning (DL) techniques. The performance of Random forests (RF), Support vector machine (SVM), Extreme gradient boosting (XGB), and K-nearest neighbors (KNN) models is evaluated, achieving overall validation accuracies of 78%, 74%, 80%, and 69%, respectively for bitou bush detection using MS imagery. Subsequently, these models are assessed on HS data, resulting in overall validation accuracies of 77%, 86%, 86%, and 80% for RF, SVM, XGB, and KNN, respectively. Moreover, the DL U-Net model achieved an overall validation accuracy of 92%, outperforming the classical ML models in MS data segmentation tasks. The results of this study highlight the superior performance of the U-Net model in comparison to classical ML models in RS data segmentation, indicating the value of DL techniques for more accurate and robust RS applications such as bitou bush detection and mapping. The insights gained from this research will aid researchers and land managers select appropriate models based on the complexity and characteristics of their RS datasets. Moreover, the integration of UAV RS and artificial intelligence (AI) provide a valuable and efficient platform for bitou bush monitoring and management practices, ultimately enhancing the efficiency and sustainability of weed control efforts.
UR - http://www.scopus.com/inward/record.url?scp=85189548140&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85189548140&partnerID=8YFLogxK
U2 - 10.1016/j.rsase.2024.101151
DO - 10.1016/j.rsase.2024.101151
M3 - Article
SN - 2352-9385
VL - 34
JO - Remote Sensing Applications: Society and Environment
JF - Remote Sensing Applications: Society and Environment
M1 - 101151
ER -