Corrected digital elevation model in coastal wetlands in Nassau and Duval Counties, Florida, 2018

High-resolution elevation data provide a foundational layer needed to understand regional hydrology and ecology under contemporary and future-predicted conditions with accelerated sea-level rise. While the development of digital elevation models (DEMs) from light detection and ranging data has enhanced the ability to observe elevation in coastal zones, the elevation error can be substantial in densely vegetated coastal wetlands. In response, we developed a machine learning model to reduce vertical error in coastal wetlands for a 1-m DEM from 2018 that covered Nassau and Duval Counties, Florida. Error was reduced by using a random forest regression model within situ observations and predictor variables from optical and radar-based satellite data and elevation derivatives. Vegetation and elevation data were collected using a real-time kinematic global positioning system (RTK GPS) in coastal wetlands at the National Park Service’s Timucuan Ecological and Historic Preserve in summer 2021 and winter 2022 (n = 344). Predictor variables included information on vegetation greenness, wetness, elevation, and vegetation structure. In the extent of coastal wetlands in Nassau and Duval Counties, the original DEM had a mean absolute error of 0.17-m and a 95th percentile error of 0.48 m. Leave-one-out cross-validation was used to assess the accuracy of the corrected DEM. In coastal wetlands, the corrected DEM had a mean absolute error of 0.08 cm and a 95th percentile error of 0.25 m. The random forest model led to a decrease in the mean absolute error by about 50% and a decrease in 95th percentile by 49%.

高分辨率高程数据可为在当代及未来预测的海平面加速上升情景下，理解区域水文与生态系统提供核心基础支撑层。尽管基于激光探测与测距（light detection and ranging）技术构建数字高程模型（digital elevation models, DEMs）的方法，已显著提升了海岸带高程观测能力，但在植被茂密的滨海湿地中，高程误差仍可能十分显著。为应对这一问题，我们针对2018年覆盖美国佛罗里达州拿骚县与杜瓦尔县的1米分辨率DEM，开发了一款机器学习模型以降低滨海湿地的垂直高程误差。本研究采用随机森林回归模型，以原位观测数据、光学与雷达卫星数据及高程衍生变量作为预测因子，实现了高程误差的修正。2021年夏季与2022年冬季，我们在美国国家公园管理局的蒂穆夸恩生态与历史保护区的滨海湿地中，采用实时动态全球定位系统（real-time kinematic global positioning system, RTK GPS）采集了植被与高程数据，共获取有效样本344组（n=344）。本次研究的预测因子涵盖植被绿度、湿度、高程及植被结构相关信息。在拿骚县与杜瓦尔县的滨海湿地范围内，原始DEM的平均绝对误差为0.17米，95分位误差达0.48米。我们采用留一交叉验证法对修正后DEM的精度进行评估，结果显示，经修正后的DEM在滨海湿地中的平均绝对误差为0.08厘米，95分位误差为0.25米。该随机森林模型使平均绝对误差降低约50%，95分位误差降低49%。