Data from: "Cross-realm transferability of species distribution models – species characteristics matter more than modelling methods applied"

Abstract This data contains occurrence observations (presence-absence) of 11 aquatic macrophytes from Bothnian Sea and Lake Puruvesi, and environmental covariates used to build species distribution models (SDMs) in paper "Cross-realm transferability of species distribution models – species characteristics matter more than modelling methods applied". At the moment the data and the code are supplied anonymously for double-blind peer review. In addition to data files, also R code for fitting the SDMs is supplied, as is the R code to replicate the analysis conducted in the paper. The data is stored in rdata format (point data), without coordinate information due to data policy restrictions. The species in the data are Isoëtes lacustris, Isoëtes echinospora, Ranunculus reptans, Ranunculus schmalhausenii, Potamogeton berchtoldii, Potamogeton perfoliatus, Potamogeton gramineus, Myriophyllum alterniflorum, Equisetum fluviatile, Eleocharis acicularis and Elodea canadensis. The environmental covariates are bottom water salinity, turbidity, sandy substrate occurrence, colored dissolved organic matter (CDOM), surface fetch, sampling depth, total nitrogen and total phosphorus, and distance to closest Phragmites australis reed bed. Objective of the study The modelling objective of the paper was species distribution model (SDM) transferability assesment. Transferability was assessed using models built in marine areas in projecting the distributions of the target species in Lake Puruvesi, Saimaa, Eastern Finland. Macrophyte mapping data from Lake Puruvesi was used as independent test data, against which transferability of the models was assessed. Location The species data was collected from two geographic areas: Bothnian Bay (Baltic Sea) and Lake Puruvesi (Eastern Finland). The marine observations from Bothnian Bay were split into three overlapping areas (areas 1-3), to test the effect of input data gradient length to SDM transferability. The largest marine sampling area (Area 3) ranged from 62.95, 65.91 latitude and 19.14, 27.99 longitude. Area 2 ranged from 63.95, 65.91 latitude and 21.53, 27.99 longitude. Area 1 ranged from 64.91, 65.91 latitude and 23.82, 27.99 longitude. The Hummonselkä subbasin of Lake Puruvesi, where the macrophyte test data was collected, is located at 61.89, 62.05 latitude and 29.58, 29.78 longitude. Species data The species observations were collected using diving transects placed in the floor of the sea or lake, and species observations were recorded in 2 m22 grid cells separated by 10 meters along the transect or 1 meters depth, depending which criteria was met first. The species data was collected in 2010 - 2020 from marine area, and 2017 from Lake Puruvesi. All macrophytes in 2 x 1 m frames were identified to species level by the diver, and the data contained information on species presence or absence in each grid cell. The diving transects were conducted using systematic survey protocol used in the underwater inventories of the Finnish Underwater Biodiversity Survey Program (VELMU). The locations of the diving transects were not randomly distributed, but were placed using expert judgement. As all our study species are macroscopic and relatively easily identifiable in the field (with the exception of possibility of mixing I. echinospora and I. lacustris), we consider the absences in our observation data to indicate true absences. That said, as the observation area is rather small (2 m2), it is possible that a species may be found in the site of investigation (e.g. a small lagoon) but be located outside the vegetation sampling grid. Environmental data Bottom water salinity The seasonal mean bottom water salinity was modeled using a generalized additive model with mean salinity as response, with log-link and gamma distribution for the errors. This was necessary to keep the resulting predictions positive. Bottom depth, CDOM, river influence and spatial location were used as predictors. Data from 448 locations were used and each location had a minimum of three observations. The model was validated using 30 % of the data left outside of the model fitting. The explained deviance of the model was 0.94 and the correlation between raw data and predicted values was 0.95 with few outliers. Turbidity Maps of turbidity (in FNU, Formazin Nephelometric Unit) were generated from Sentinel-2 Multi-Spectral Imager (MSI) observations using the Case-2 Regional Coast Colour (C2RCC) bio-optical inversion model, containing separate atmospheric correction and water quality parts. Before computing C2RCC, the original 10-meter input data was downsampled to 60 meters. The output variable of the C2RCC processor correlative to turbidity is the backscattering of total suspended sediments at 443 nm, which was further calibrated into turbidity (FNU) values using SYKE's empirical equations for coastal waters and clear lakes (for a similar approach, see Attila et al. (2013) and Sagerman, Hansen, and Wikström (2020)). Monthly observations of turbidity were aggregated into median composites to reduce the effects of cloud cover and other disturbances. Due to low solar elevation and ice cover in winter, the turbidity distribution maps are generated only for the summer months (May to September). The current processing covers years 2017 to 2021. An average raster layer was created from monthly observations as input for SDM building. Probability of sandy substrate Random forest model was used to classify sandy bottoms from Sentinel 2 MSI satellite images in shallow water areas. Identifying sandy substrate is based on the higher reflectance compared to other substrates. The model was trained and validated using diver recorded field observations in the Baltic area, and diver recorded and echo sounding observations in the freshwater area. For full coverage including areas beyond the shallow water, the satellite image classification was combined with boosted regression tree modelling result in the Baltic, and echo sounding based product in the freshwater region. The resulting layers were probabilities of sandy substrate with 10-meter cell resolution. CDOM We used different methods to estimate the CDOM levels in Bothnian Bay and Lake Puruvesi, based on biogeochemical model data and satellite images. For Lake Puruvesi, we applied the Finnish Environment Institute's (Syke) in-house CDOM algorithm to the Sentinel-2 MSI images processed by the C2RCC bio-optical processor (Brockmann et al. 2016). The observations in 10 m resolution were aggregated as monthly averages for each month of the summer season (May to October) from 2017 to 2021. For Bothnian Bay, we used Syke's in-house Sentinel-2 MSI CDOM layers (resolution: 60 m) aggregated as seasonal averages (1 Jul to 7 Sep). CDOM values are given as absorption coefficient of CDOM at 400 nm [m⁻¹]. Surface fetch A surface fetch raster was produced to the Puruvesi and Bothnian Bay. The analysis required a feature layer of shorelines from Puruvesi and Baltic Sea. First, we created polyline from north to south spanning over the whole area of interest with a gap of 20 meters which is also the resolution of the output raster. These lines were then cut each time they hit the shoreline and the part of the line that was overlapping land was removed. The distance of the remaining lines was then calculated and a point with the distance value was created every 20 meters. Each time the line was cut when hitting an island for example and starting again from the other side of the island, the distance calculation started from 0. This created a point dataset with a distance value in each point. We repeated the procedure for 15 times for different compass directions with 22.5 degree intervals and calculated average fetch for each point location on 20 meters grid from these 15 point layers. Depth Depth was measured by a diver using a dive computer while surveying each vegetation grid cell, and measured depth was used when projecting model results to Puruvesi (transferability performance). In addition, a depth model for the freshwater region was created from Sentinel 2 MSI satellite image using the logarithmic band ratio model of blue and red band. The model was calibrated using diver recorded field observations and validated against echo sounding measurements. For more complete coverage and to include deep areas, echo soundings from multiple sources were combined with the satellite derived bathymetry. The cell resolution of the resulting depth layer was 10 meters. Total nitrogen and phosphorus Mean total nitrogen and phosphorus layers for marine area were produced using ArcGIS "splines with barriers" tool for the EEZ of Finland with 20 meters spatial resolution (Virtanen et al. 2018). Summer (July - September) nutrient measurements from 0 to 10 meters depth between 2010 and 2020, obtained from the VESLA database, were used as input data for the interpolation. Nitrogen and phosphorus measurements in Puruvesi between 2010 and 2020 was gathered from the VESLA database. Data from July to September was selected to represent the growing season. A mean value of NTOT and PTOT was then calculated for each location. Spline with Barriers (SwB) tool was used to interpolate the values (Arcmap 10.7.1). The tool uses a feature layer as barrier to create the raster representing only the area of interest. For the barrier and the extent of the interpolated raster we used a shapefile representing Lake Puruvesi shoreline. The resolution was set to 5x5 meters. SwB tool created an "extent box" around the area of interest which was removed with Extract by Mask tool using the shoreline feature layer. After the interpolation we noticed that either one of the locations was situated on land or the polygon used as barrier was "leaking". SwB doesn´t interpolate areas that doesn't have locations with values or aren´t connected to the main body of water. To fix this, the raster was extended outwards based on the values of nearby cells and after that the raster was masked again to remove any cells on land. The phosphorus interpolation provided negative values in southern parts on Enanlahti in Kontiolahti and Muholanlahti. These negative values were caused by considerably larger phosphorus values in Enanlahti Lamminniemi (9m) Enanlahti Lamminniemi (4m) locations when compared with the nearby Puruvesi Enanlahti location. The interpolation apparently continued to decrease the values according to the trend set by the difference between these locations and caused it to reach negative values. The southern parts of the bay, about 750 meters, was removed and new values were calculated based on the surrounding cells with Focal Statistics tool. The interpolations were validated by removing 20 % of the locations and reproducing the interpolation. The removed locations and their values were then compared to the interpolated raster. R∗2∗2 value from phosphorus interpolation model was 0.91 after removing two outliers and R22 value from nitrogen interpolation model was 0.715 after removing one outlier. Distance to closest reed The aquatic vegetation (Phragmites australis reeds) presence/absence maps were also generated from Sentinel-2 MSI data. The processing included extracting one month of data (July 2019) from green and near-infra-red bands from Sentinel-2 Global Mosaic (S2GM) service and transforming those to normalized-difference vegetation indices (NDVIs). After that, Bayesian statistics were used to predict the posterior probability of vegetation occurrence when distance from shore and NDVI were used as predictor variables. The posterior variable was thresholded and the resulting vegetation presence areas were sieved so that both too small vegetation areas (fewer than 5 pixels) or areas that were not directly attached to shoreline were removed. The resulting map has 10 m pixel size and tentatively represents the locations of reed belts or other shoreline-attached vegetation. This EO-based layer could also be referred to as helophytes or helophytic macrophytes, as it denotes a specific zone of vegetation with emergent aquatic plants containing leaf-green, particularly those that grow densely and have horizontally oriented leaves. In some lakes, this layer can represent, for example, thick stands of Equisetum fluviatile, although in most cases, it is associated with common reed belts. The approach is described in more detail in Koponen et al. (2022). Data partitioning Data was partitioned with 70/30 splitting into training and test (interpolation accuracy) data. The splitting was repeated 100 times for each species by randomly selecting 70 % of observations which were used to build each of the SDMs (GLM, GAM, BRT and BART). The partitioning was repeated for each of the three input data areas and 11 species. The input data indexes for replicating the split are supplied in the data files. R code Code files contain scripts for fitting the SDM models described in the paper using the data. Also code for beta regression analyses for the analysis of the modelling results conducted in the paper, are supplied.**摘要** 本数据集包含采自波的尼亚湾与普鲁韦西湖的11种大型水生植物的出现观测数据（存在-缺失数据），以及论文《物种分布模型的跨域可迁移性——物种特征比建模方法更为重要》中用于构建物种分布模型（species distribution models, SDMs）的环境协变量。目前为配合双盲同行评议，本数据集与代码以匿名形式提供。除数据文件外，本包还提供了用于拟合SDMs的R代码，以及复现论文中所有分析流程的R代码。由于数据政策限制，本数据以rdata格式（点位数据）存储，未包含坐标信息。 本数据包含的物种为湖生水韭（Isoëtes lacustris）、刺孢水韭（Isoëtes echinospora）、匍枝毛茛（Ranunculus reptans）、施马尔豪森毛茛（Ranunculus schmalhausenii）、贝氏眼子菜（Potamogeton berchtoldii）、穿叶眼子菜（Potamogeton perfoliatus）、禾叶眼子菜（Potamogeton gramineus）、异花狐尾藻（Myriophyllum alterniflorum）、水生木贼（Equisetum fluviatile）、针蔺（Eleocharis acicularis）以及加拿大伊乐藻（Elodea canadensis）。所用环境协变量包括底层水盐度、浊度、沙质底质出现概率、有色溶解有机物（colored dissolved organic matter, CDOM）、水面风区长度、采样水深、总氮与总磷，以及至最近的芦苇（Phragmites australis）群落的距离。 **研究目标** 本论文的建模目标为开展物种分布模型（SDMs）的可迁移性评估。评估方式为：使用在海洋区域构建的模型，对芬兰东部塞马湖普鲁韦西湖中的目标物种分布进行预测，并以普鲁韦西湖的大型水生植物调查数据作为独立测试数据，以此评估模型的可迁移性。 **研究区域** 物种数据采自两个地理区域：波的尼亚湾（波罗的海）与芬兰东部的普鲁韦西湖。为检验输入数据梯度长度对SDM可迁移性的影响，我们将波的尼亚湾的海洋观测数据划分为3个重叠区域（区域1-3）。最大的海洋采样区域（区域3）的纬度范围为62.95°N~65.91°N，经度范围为19.14°E~27.99°E；区域2的纬度范围为63.95°N~65.91°N，经度范围为21.53°E~27.99°E；区域1的纬度范围为64.91°N~65.91°N，经度范围为23.82°E~27.99°E。本研究的大型水生植物测试数据采自普鲁韦西湖的Hummonselkä子流域，该区域的纬度范围为61.89°N~62.05°N，经度范围为29.58°E~29.78°E。 **物种数据** 本研究通过布设至海底或湖底的潜水样带开展物种观测，将样带内的区域划分为2平方米的网格单元，网格单元间距沿样线方向为10米，或按水深间隔1米，以先满足的条件为准。海洋区域的物种数据采集于2010-2020年，普鲁韦西湖的物种数据采集于2017年。潜水员在2×1米的样框内对所有大型水生植物进行物种鉴定，并记录每个网格单元内物种的存在或缺失情况。本研究的潜水样带采用芬兰水下生物多样性调查项目（Finnish Underwater Biodiversity Survey Program, VELMU）的水下系统调查规程开展。潜水样带的布设并非随机，而是基于专家判断确定点位。本研究的所有目标物种均为大型水生植物，野外可较为便捷地识别（仅存在刺孢水韭与湖生水韭混淆的可能），因此我们将观测数据中的物种缺失记录视为真实的物种缺失。不过由于观测单元面积较小（2平方米），有可能在调查点位（如小型潟湖）中存在目标物种，但该物种并未被纳入植被采样网格的情况。 **环境数据** **底层水盐度** 我们以底层水平均盐度为响应变量，采用对数连接函数与伽马误差分布的广义可加模型（generalized additive model, GAM）对季节平均底层水盐度进行建模，以确保预测结果为正值。模型的预测因子包括水深、CDOM、河流影响程度与空间位置。本研究使用了448个点位的观测数据，每个点位至少包含3次观测。我们将30%的建模数据预留作为验证集，以验证模型性能。该模型的解释偏差为0.94，原始观测值与预测值的相关系数为0.95，仅存在少量异常值。 **浊度** 本研究基于Sentinel-2多光谱成像仪（Multi-Spectral Imager, MSI）的观测数据，采用Case-2区域海岸颜色（Case-2 Regional Coast Colour, C2RCC）生物光学反演模型生成浊度分布图（单位为福尔马肼浊度单位，Formazin Nephelometric Unit, FNU），该模型包含独立的大气校正与水质反演模块。在运行C2RCC模型前，我们将原始10米分辨率的输入数据重采样至60米分辨率。C2RCC处理器输出的与浊度相关的变量为443 nm处总悬浮颗粒物的后向散射系数，我们随后使用芬兰环境研究所（SYKE）针对近岸水域与清澈湖泊的经验公式，将该系数校准为浊度（FNU）值（类似方法可参考Attila等(2013)与Sagerman、Hansen及Wikström(2020)的研究）。为减少云覆盖与其他干扰的影响，我们将月度浊度观测数据聚合为中值合成数据集。由于冬季太阳高度角较低且存在冰覆盖，浊度分布图仅生成夏季（5-9月）的结果。当前的处理覆盖2017-2021年的数据，我们从月度观测数据中生成平均栅格图层，作为构建SDMs的输入数据。 **沙质底质出现概率** 我们使用随机森林模型，基于Sentinel-2 MSI卫星影像对浅水区的沙质底质进行分类。沙质底质的识别基于其相较于其他底质更高的反射率。模型的训练与验证采用波罗的海区域的潜水实地观测数据，以及淡水区域的潜水观测与回声测深数据。为实现全区域覆盖（包括浅水区以外的区域），我们将卫星图像分类结果与波罗的海区域的提升回归树建模结果、淡水区域的回声测深产品进行融合。最终生成的沙质底质出现概率栅格图层分辨率为10米。 **有色溶解有机物（CDOM）** 我们基于生物地球化学模型数据与卫星影像，采用不同方法分别估算波的尼亚湾与普鲁韦西湖的CDOM浓度。对于普鲁韦西湖，我们将芬兰环境研究所（SYKE）的内部CDOM算法应用于经C2RCC生物光学处理器处理的Sentinel-2 MSI影像（Brockmann等, 2016）。我们将10米分辨率的观测数据聚合为2017-2021年夏季（5-10月）各月的月均数据。对于波的尼亚湾，我们使用SYKE内部的Sentinel-2 MSI CDOM图层（分辨率：60米），并聚合为季节平均数据（7月1日-9月7日）。CDOM值以400 nm处CDOM的吸收系数[m⁻¹]表示。 **水面风区长度** 我们为普鲁韦西湖与波的尼亚湾生成了水面风区长度栅格图层。本分析需要普鲁韦西湖与波罗的海的岸线矢量图层。首先，我们在研究区域内生成南北走向的折线，折线间距为20米（与输出栅格分辨率一致）。随后，每当折线与岸线相交时，我们将该折线截断，并移除与陆地重叠的线段。接着，我们计算剩余线段的长度，并在每20米的间隔处生成一个带有距离值的点位。若折线因穿过岛屿等障碍物被截断，则从岛屿另一侧重新开始，此时距离计算将重置为0。该流程最终生成一个每个点位均带有距离值的点数据集。我们以22.5°为间隔，针对15个不同的方位角重复上述流程，并基于这15个点图层，计算20米分辨率栅格上每个点位的平均风区长度。 **水深** 在对每个植被网格单元进行调查时，潜水员使用潜水电脑记录水深，该实测水深数据用于将模型结果投影至普鲁韦西湖（以评估可迁移性性能）。此外，我们基于Sentinel-2 MSI卫星影像，利用蓝波段与红波段的对数波段比模型构建了淡水区域的水深模型。模型的校准采用潜水实地观测数据，验证则采用回声测深数据。为实现更完整的区域覆盖并纳入深水区，我们将多源回声测深数据与卫星反演的水深数据进行融合。最终生成的水深栅格图层分辨率为10米。 **总氮与总磷** 我们使用ArcGIS的“带障碍样条插值”工具（splines with barriers），以20米的空间分辨率生成了芬兰专属经济区的海洋区域总氮与总磷平均图层（Virtanen等, 2018）。插值的输入数据为2010-2020年采自0-10米水深的夏季（7-9月）营养盐观测数据，数据来源于VESLA数据库。普鲁韦西湖2010-2020年的氮、磷观测数据同样来源于VESLA数据库，我们选取7-9月的数据代表生长季。我们为每个点位计算总氮（NTOT）与总磷（PTOT）的平均值，并采用带障碍样条插值（Spline with Barriers, SwB）工具进行空间插值（使用ArcMap 10.7.1）。该工具通过矢量障碍图层仅生成研究区域范围内的栅格数据，我们使用普鲁韦西湖岸线的shapefile作为障碍图层与插值栅格的范围，分辨率设置为5×5米。SwB工具会自动生成研究区域的“范围框”，我们随后使用岸线矢量图层通过“按蒙版提取”工具移除该范围框。插值完成后，我们发现部分点位位于陆地区域，或障碍多边形存在“泄漏”问题。SwB工具不会对无观测值或未与主水体连通的区域进行插值。为解决该问题，我们基于邻近单元格的值将栅格向外扩展，随后再次进行蒙版处理以移除陆地区域的单元格。在孔蒂奥拉赫蒂的埃南拉赫蒂湾南部与穆霍兰拉赫蒂湾，总磷插值结果出现了负值。该负值的产生原因是埃南拉赫蒂的Lamminniemi（9m水深）与Lamminniemi（4m水深）点位的磷浓度显著高于附近的普鲁韦西湖埃南拉赫蒂点位，插值过程按照该点位间的差异趋势持续降低数值，最终导致出现负值。我们移除了该海湾南部约750米的区域，并使用焦点统计工具基于周围单元格计算新的数值。我们通过预留20%的观测点位并重新执行插值流程来验证插值结果，将预留点位的实测值与插值栅格进行对比。移除2个异常值后，总磷插值模型的决定系数R²为0.91；移除1个异常值后，总氮插值模型的决定系数R²为0.715。 **至最近芦苇群落的距离** 我们同样基于Sentinel-2 MSI数据生成了水生植被（芦苇，Phragmites australis）的存在-缺失分布图。处理流程包括从Sentinel-2全球镶嵌数据集（Sentinel-2 Global Mosaic, S2GM）中提取2019年7月的绿光与近红外波段数据，并将其转换为归一化植被指数（normalized difference vegetation index, NDVI）。随后，我们以距岸距离与NDVI作为预测变量，采用贝叶斯统计方法预测植被出现的后验概率。对后验变量进行阈值划分后，我们对生成的植被存在区域进行筛选，移除面积小于5个像素的区域以及未直接与岸线相连的区域。最终生成的植被分布图分辨率为10米，初步代表了芦苇带或其他近岸附着型植被的分布位置。该基于遥感（EO）的图层也可称为挺水植物（helophytes）或挺水型大型水生植物图层，因为其表征的是一类具有绿叶的挺水水生植物群落，尤其是那些生长密集、叶片水平分布的物种。在部分湖泊中，该图层也可代表茂密的水生木贼（Equisetum fluviatile）群落，但在大多数情况下，其对应的是普通芦苇带。该方法的详细描述可参考Koponen等(2022)的研究。 **数据划分** 我们将数据以7:3的比例划分为训练集与测试集（用于插值精度评估）。针对每个物种，我们随机选取70%的观测数据用于构建物种分布模型（包括广义线性模型GLM、广义可加模型GAM、提升回归树BRT与贝叶斯加法回归树BART），并重复该划分流程100次。该划分流程针对3个输入数据区域与11个物种分别执行。数据文件中提供了用于复现数据划分的输入数据索引。 **R代码** 代码文件包含了使用本数据集拟合论文中所述SDM模型的脚本，同时提供了用于分析论文建模结果的贝塔回归分析代码。