Evaluating Normalization Methods for Seagrass Mapping with Supervised Classification of PlanetScope Imagery
College
College of Sciences
Department
Ocean and Earth Sciences
Graduate Level
Doctoral
Graduate Program/Concentration
Oceanography
Presentation Type
Poster Presentation
Abstract
Accurate mapping of submerged aquatic vegetation (SAV) is crucial for monitoring coastal ecosystems. However, this process is complicated by the fact submerged objects exhibit low reflectance due to significant absorption and scattering of light within the water column, resulting in a dark appearance. In contrast, terrestrial surfaces generally have higher reflectance, as they are not subject to the attenuating effects of water and instead reflect a greater proportion of incident sunlight. This imbalance creates a limited dynamic range among water pixels, complicating classification of submerged targets. Normalization of reflectance values in each spectral band may mitigate these issues by increasing dynamic range across submerged targets and equalizing values across the time-series of imagery, improving classification accuracy. This study evaluates the effectiveness of two normalization methods in improving the supervised classification of Planet SuperDove 8-band imagery for SAV detection in Pocomoke Sound, a mesohaline region of the Chesapeake Bay. Analysis focuses on summer imagery from 2020 to 2024, aligning with the peak seagrass growing season. To assess the impact of normalization, the frequency distribution of pixel values was examined in each band before and after applying min-max normalization and 1st/99th percentile normalization, first masking out land and bright areas to enhance contrast, then normalizing the remaining dark pixels to improve classification performance. In addition to normalization, the study involved testing two indices—OSW (Optically Shallow Water Index based on the green, red, and near-infrared bands) and Hue (represents spectral variation as a single color value instead of three RGB channels) —to determine whether they enhance classification accuracy by better distinguishing SAV from surrounding water and other features. Effective normalization has the potential to enable trained classification models to be applied to future images of the same site without requiring additional training data. The findings from this study may provide valuable insights into preprocessing strategies for deep-learning-based SAV mapping, contributing to the development of automated and scalable methods for coastal ecosystem monitoring.
Keywords
Seagrass, Vegetation indices, Seagrass mapping
Evaluating Normalization Methods for Seagrass Mapping with Supervised Classification of PlanetScope Imagery
Accurate mapping of submerged aquatic vegetation (SAV) is crucial for monitoring coastal ecosystems. However, this process is complicated by the fact submerged objects exhibit low reflectance due to significant absorption and scattering of light within the water column, resulting in a dark appearance. In contrast, terrestrial surfaces generally have higher reflectance, as they are not subject to the attenuating effects of water and instead reflect a greater proportion of incident sunlight. This imbalance creates a limited dynamic range among water pixels, complicating classification of submerged targets. Normalization of reflectance values in each spectral band may mitigate these issues by increasing dynamic range across submerged targets and equalizing values across the time-series of imagery, improving classification accuracy. This study evaluates the effectiveness of two normalization methods in improving the supervised classification of Planet SuperDove 8-band imagery for SAV detection in Pocomoke Sound, a mesohaline region of the Chesapeake Bay. Analysis focuses on summer imagery from 2020 to 2024, aligning with the peak seagrass growing season. To assess the impact of normalization, the frequency distribution of pixel values was examined in each band before and after applying min-max normalization and 1st/99th percentile normalization, first masking out land and bright areas to enhance contrast, then normalizing the remaining dark pixels to improve classification performance. In addition to normalization, the study involved testing two indices—OSW (Optically Shallow Water Index based on the green, red, and near-infrared bands) and Hue (represents spectral variation as a single color value instead of three RGB channels) —to determine whether they enhance classification accuracy by better distinguishing SAV from surrounding water and other features. Effective normalization has the potential to enable trained classification models to be applied to future images of the same site without requiring additional training data. The findings from this study may provide valuable insights into preprocessing strategies for deep-learning-based SAV mapping, contributing to the development of automated and scalable methods for coastal ecosystem monitoring.