ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Image-based Assessment of Vegetation Cover and Composition using U-Net-based Semantic Segmentation

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2023 ASABE Annual International Meeting  2300237.(doi:10.13031/aim.202300237)
Authors:   Puranjit Singh, Yin Bao, Michael A. Perez, Wesley N. Donald
Keywords:   Deep learning, Semantic segmentation, UNet, Vegetation prediction, Multi-class classification, Grasses, Intersection over Union, Coefficient of determination – R2.

Abstract. The vegetation cover and composition assessment is critical to understanding ecosystem health and biodiversity conservation. In the recent past, several factors such as rapid urbanization, soil degradation, and climate change have led to significant alterations in vegetation cover and composition, and have compromised Earth‘s natural ecosystem. Effective vegetation cover and composition management are quintessential to the sustainable growth of our planet. The field of vegetation cover and composition assessment is expanding as it realizes the essentiality of conserving biodiversity, effective land use planning, and optimum resource management strategies at our disposal. The recent advancements in deep learning and remote sensing technologies enable the collection of spatial information from the Earth‘s surface and the extraction of useful insights for informed decision-making that could benefit the environment. The aim of this research was to develop advanced semantic segmentation-based deep learning models using convolutional neural networks (CNNs) that can effectively identify and differentiate multiple vegetation species, and accurately predict the area under vegetation cover. To achieve this target a dataset curation was done which comprised 10 diverse vegetation species especially ‘grasses‘. The deep learning model of UNet was trained using different architecture backbones to generate composition assessment amongst different grass species. Also, trained predictive models were also used to evaluate the area under vegetation cover as well. The best-performing model was able to achieve an R2 value of 0.84 in predicting the vegetation cover and a mean IOU score of 0.54 during testing. This study demonstrates a potential to use remote sensing data to generate vegetation cover and composition assessment for farming, urban planning, and construction applications.

(Download PDF)    (Export to EndNotes)