Spatially Quantifying Forest Damage from Hurricane Michael using Sentinel-2 Imagery

<strong>ABSTRACT:</strong> Hurricane Michael made landfall on Mexico Beach, Florida panhandle as a Category 5 storm on October 10^th, 2018. The storm had a large impact on the forests in the Florida panhandle and into Georgia. In this study we use Sentinel-2 imagery and 248 forest plots collected prior to landfall in 2018 in the forests impacted by Hurricane Michael to build a general linear model of tree basal area across the landscape. The basal area model was constrained to areas where trees were present using a tree presence model as a hurdle. We informed the model with post hurricane Sentinel-2 imagery and compared the pre and post hurricane basal area maps to assess the loss of basal area following the hurricane. The basal area model had an r-squared value of 0.508. Our results provide a detailed map showing the extent of basal area loss across the Florida panhandle at 10m spatial scale. Plots were revisited to ground truth the modelled results and showed that the model performed well at categorizing forest hurricane damage. This study demonstrates the use of remotely sensed imagery and in-situ forest measurements to rapidly quantify, using common forestry metrics, forest damage from large natural disturbances at spatial resolution useful to inform disaster response management decisions. <strong>METHODS:</strong> The Restore .csv file is data from forestry plots established by the Florida Natural Areas Inventory (FNAI) as a baseline for a Restore Act project focused on the Florida panhandle. A total of 248 plots were visited between December 2017 and March 2018. These temporary plots were navigated to using handheld GPS units and laid out in 36m squares containing four 9m diameter non overlapping subplots. Measurements of vegetative cover, tree species count, tree condition, and diameter at breast height were taken for all trees in the subplots. Post hurricane Michael 70 plots were revisited, measurements at these plots were a subjective plot hurricane damage categorization, and count and diameter of downed or damaged trees as well as miscellaneous notes regarding site damage. The state parks .csv file is data from forestry plots established by the Florida Natural Areas Inventory (FNAI) at Florida State parks post hurricane Michael. These plots are 20 m circular radius that include subjective plot hurricane damage categorization, and count of downed or damaged trees, herbaceous cover, as well as miscellaneous notes regarding site damage. Several fields were added to these plot .csv files post field visit as a variables extracted from a principal component analysis that used Sentinel-2 imagery to inform a remote sensing analysis of basal area. A file geodatabase is attached that contains the project area boundary, Apalachicola National Forest boundary, the three shapefiles of all restore plots, revisited restore plots, and state parks plots. Raster outputs from our analysis are also available in this gdb, they contain metadata in their item descriptions. The general metadata for these rasters follows: A general linear regression model was built to estimate tree basal area across the study area of 11 counties in the Florida Panhandle. Basal area (BA) was calculated from Restore field plots where trees were present located in and around the Apalachicola National Forest. Plot measurements include all trees within four non-overlapping 9m radius circular subplots within a 36m square plot. Tree diameter at breast height (DBH), species, count, and condition measurements were recorded. Measurements were summarized to the plot and DBH (square inches) was converted to basal area per acre (square feet per acre) using the formula 0.005454 * DBH^2. Plot measurements of basal area per acre were related to the top ten principal components from a principal component analysis (PCA) at the spatial resolution of the Restore plots (40 m). The PCA used the normalized Sentinel-2 spectral values and two texture values from the 7 mosaicked images from two time periods, the winter of 2017-2018 and the spring of 2018, before Hurricane Michael. A softmax neural network model was built from the PCA and Restore plot datasets to identify areas where trees were present. The basal area model was applied to the pre and post hurricane PCA imagery to create modelled surfaces of estimated Basal Area in pixels that where over 50% likely to contain trees according to the softmax neural network model. The basal are linear regression model results and predictors for the model are in the tables below. Table Linear Regression Model. Model fit and predictors for BAA. <strong>N</strong> <strong>RMSE</strong> <strong>R2</strong> <strong>Adjusted R2</strong> 231 2.811 0.51 0.50 Road, water and urban areas were masked out of this raster dataset using a 1 m landcover product created by the author and located here - https://doi.org/10.2737/RDS-2017-0014**摘要：** 2018年10月10日，迈克尔飓风（Hurricane Michael）以五级飓风强度在佛罗里达狭长地带的墨西哥海滩登陆。该风暴对佛罗里达狭长地带及佐治亚州的森林造成了严重影响。本研究利用哨兵二号影像（Sentinel-2）以及2018年飓风登陆前在受迈克尔飓风影响的森林中布设的248个森林样地（forest plots）数据，构建了研究区范围内树木断面积（tree basal area）的广义线性模型（General Linear Model）。该断面积模型通过 hurdle模型（Hurdle Model）约束至有树木分布的区域。研究使用飓风过后的哨兵二号影像训练模型，并对比飓风前后的断面积分布图以评估飓风后的断面积损失。本断面积模型的决定系数（R-squared）为0.508。研究结果生成了10米空间分辨率下佛罗里达狭长地带断面积损失范围的详细分布图。研究团队对部分样地进行了重访以实地验证模型结果，结果显示该模型在分类森林飓风破坏程度方面表现良好。本研究证明了结合遥感影像（Remote Sensed Imagery）与原位森林测量（In-situ Forest Measurements）数据，可利用通用林业指标快速量化大型自然扰动引发的森林破坏，其空间分辨率可为灾害响应管理决策提供有效支撑。 **研究方法：** 本研究使用的`Restore.csv`数据集源自佛罗里达自然区域名录（Florida Natural Areas Inventory, FNAI）为聚焦佛罗里达狭长地带的《恢复法案》（Restore Act）项目建立的森林样地基线数据。2017年12月至2018年3月间，研究人员共走访了248个临时样地，通过手持GPS设备导航定位，样地以36米见方的正方形布设，内含4个直径9米的不重叠子样地。对每个子样地内的所有树木开展了植被覆盖度、树木种类数量、树木健康状况以及胸径（Diameter at Breast Height, DBH）的测量。飓风迈克尔过后，研究团队重访了其中70个样地，记录了主观的样地飓风破坏等级、倒伏或受损树木的数量与胸径，以及样地破坏情况的其他备注信息。 本研究使用的`state parks.csv`数据集源自佛罗里达自然区域名录在迈克尔飓风过后于佛罗里达州立公园布设的森林样地数据。此类样地为半径20米的圆形区域，包含主观的样地飓风破坏等级、倒伏或受损树木数量、草本覆盖度，以及样地破坏情况的其他备注信息。野外调查结束后，研究人员向这些样地数据集添加了多个变量字段，这些变量源自主成分分析（Principal Component Analysis, PCA），该分析利用哨兵二号影像开展了断面积的遥感反演研究。 本次研究附带了一个文件地理数据库（File Geodatabase），其中包含研究区边界、阿巴拉契科拉国家森林边界，以及所有恢复样地、重访恢复样地和州立公园样地的3个形状文件（Shapefiles）。该地理数据库中还包含本次分析生成的栅格输出结果（Raster Outputs），其元数据已记录在项描述中。这些栅格数据的通用元数据如下：本研究构建了广义线性回归模型，用于估算佛罗里达狭长地带11个县的研究区范围内的树木断面积。断面积（Basal Area, BA）的计算基于阿巴拉契科拉国家森林及其周边区域内有树木分布的恢复样地实地测量数据。每个样地内设置4个不重叠的半径9米圆形子样地，对其中所有树木开展测量，记录胸径、树种、数量及健康状况。将样地水平的测量数据汇总后，利用公式`0.005454 * DBH²`将以平方英寸计的胸径转换为每英亩断面积（单位：平方英尺/英亩）。将每英亩断面积的样地测量值与以恢复样地空间分辨率（40米）计算得到的主成分分析前10个主成分进行关联。该主成分分析使用了哨兵二号光谱归一化值，以及2017-2018年冬季与2018年春季两个时段拼接的7景影像的2个纹理特征值，数据采集于迈克尔飓风登陆前。 研究团队基于主成分分析与恢复样地数据集构建了Softmax神经网络（Softmax Neural Network）模型，用于识别有树木分布的区域。将断面积模型应用于飓风前后的主成分分析影像，生成了基于像素的断面积估算表面，仅保留Softmax神经网络模型判定为有50%以上概率存在树木的像素区域。本断面积线性回归模型的结果及模型预测因子见下表。 表 线性回归模型：断面积（BAA）的模型拟合度与预测因子 | 样本量（N） | 均方根误差（Root Mean Square Error, RMSE） | 决定系数（R²） | 调整后决定系数（Adjusted R²） | |------------|-------------------------------------------|---------------|--------------------------------| | 231 | 2.811 | 0.51 | 0.50 | 研究人员利用作者制作的1米分辨率土地覆盖产品，将道路、水体与城市区域从该栅格数据集中掩膜去除，该产品的链接为：https://doi.org/10.2737/RDS-2017-0014