Department of Geosciences, INRS-ETE, University of Québec, Québec, Canada.
PLoS One. 2011;6(6):e21265. doi: 10.1371/journal.pone.0021265. Epub 2011 Jun 20.
Epi-macrobenthic species richness, abundance and composition are linked with type, assemblage and structural complexity of seabed habitat within coastal ecosystems. However, the evaluation of these habitats is highly hindered by limitations related to both waterborne surveys (slow acquisition, shallow water and low reactivity) and water clarity (turbid for most coastal areas). Substratum type/diversity and bathymetric features were elucidated using a supervised method applied to airborne bathymetric LiDAR waveforms over Saint-Siméon-Bonaventure's nearshore area (Gulf of Saint-Lawrence, Québec, Canada). High-resolution underwater photographs were taken at three hundred stations across an 8-km(2) study area. Seven models based upon state-of-the-art machine learning techniques such as Naïve Bayes, Regression Tree, Classification Tree, C 4.5, Random Forest, Support Vector Machine, and CN2 learners were tested for predicting eight epi-macrobenthic species diversity metrics as a function of the class number. The Random Forest outperformed other models with a three-discretized Simpson index applied to epi-macrobenthic communities, explaining 69% (Classification Accuracy) of its variability by mean bathymetry, time range and skewness derived from the LiDAR waveform. Corroborating marine ecological theory, areas with low Simpson epi-macrobenthic diversity responded to low water depths, high skewness and time range, whereas higher Simpson diversity relied upon deeper bottoms (correlated with stronger hydrodynamics) and low skewness and time range. The degree of species heterogeneity was therefore positively linked with the degree of the structural complexity of the benthic cover. This work underpins that fully exploited bathymetric LiDAR (not only bathymetrically derived by-products), coupled with proficient machine learner, is able to rapidly predict habitat characteristics at a spatial resolution relevant to epi-macrobenthos diversity, ranging from clear to turbid waters. This method might serve both to nurture marine ecological theory and to manage areas with high species heterogeneity where navigation is hazardous and water clarity opaque to passive optical sensors.
底栖生物多样性、丰度和组成与沿海生态系统中海底栖息地的类型、组合和结构复杂性有关。然而，这些栖息地的评估受到水基调查（获取速度慢、水深浅、反应性低）和水清晰度（大多数沿海地区浑浊）的限制。使用应用于圣西蒙-博纳旺蒂尔近岸地区（加拿大圣劳伦斯湾魁北克）机载测深激光雷达波形的监督方法阐明了基质类型/多样性和水深特征。在 8 平方公里的研究区域内，在 300 个站点拍摄了高分辨率水下照片。基于最先进的机器学习技术（如朴素贝叶斯、回归树、分类树、C4.5、随机森林、支持向量机和 CN2 学习者），测试了七个模型，以预测作为类数函数的八个底栖生物多样性指标。随机森林在应用于底栖生物群落的三离散辛普森指数方面优于其他模型，通过平均水深、从激光雷达波形派生的时间范围和偏度解释了其变异性的 69%（分类准确性）。与海洋生态理论一致，辛普森底栖生物多样性低的区域对应于水深低、偏度和时间范围高，而辛普森多样性高的区域则依赖于更深的底部（与更强的水动力相关）和低偏度和时间范围。因此，物种异质性程度与底栖生物覆盖物的结构复杂性程度呈正相关。这项工作证明，充分利用测深激光雷达（不仅是从测深衍生的副产品），再加上熟练的机器学习，能够以从清澈到浑浊的水域都相关的空间分辨率快速预测栖息地特征。这种方法既可以促进海洋生态理论，也可以管理物种异质性高的区域，在这些区域中，导航是危险的，水清晰度对无源光学传感器是不透明的。
