Abstract:

This study mapped land cover (water, bare ground, forest, grass, marsh, algal flat, building, bridge culvert, and agriculture) around Matagorda Bay, Texas. The study area was defined by a 2-km buffer around the West Matagorda Bay shoreline and extended from the western portion of the Colorado River Delta through the eastern portion of Matagorda Island, Texas. This study incorporated WorldView-2 (WV-2; acquired on 2012-11-17, 2013-05-05, and 2013-12-16) and lidar (acquired 2018-01-04 - 2018-02-23 and 2019-01-24 – 2019-01-29) to obtain a 2-m resolution habitat map for the entire study area. A novel stacked classification approach was developed to take advantage of high-resolution satellite imagery and airborne lidar point clouds. Ultimately, a rule-based classifier was stacked on a group of machine learning classifiers for multispectral images and a filter classifier for lidar point clouds. The data were created for the Texas Office of the Comptroller project titled “Matagorda Bay Ecosystem Assessment.” Maps of vegetation, sand, and water coverage for discrete dates from 1850 to 2020 are available in related dataset HI.x833.000:0020 (https://doi.org/10.7266/zs2f74bj).

Suggested Citation:

Su, Lihong, Jessica Magolan, and James Gibeaut. 2024. High resolution habitat map of West Matagorda Bay, Texas, derived from WorldView-2 satellite imagery and lidar data, 2012-2019. Distributed by: GRIIDC, Harte Research Institute, Texas A&M University–Corpus Christi. doi:10.7266/ex6xqek7

Purpose:

This dataset was developed within the Harte Research Institute for Gulf of Mexico Studies (HRI) for the project titled "Matagorda Bay Ecosystem Assessment" in order to obtain a high resolution, detailed habitat map for West Matagorda Bay and the surrounding area.

Data Parameters and Units:

Area_SqKm - area of land cover in square kilometers Shape_Length - perimeter of feature in meters Shape_Area - area of feature in square meters

Methods:

An overview of the methods is presented below; however a more detailed methodology can be found in the report submitted to the Texas Office of the Comptroller for the project titled “Matagorda Bay Ecosystem Assessment”. A novel stacked classification approach was developed to take advantage of high-resolution satellite imagery and airborne lidar point clouds. Ultimately, a rule-based classifier was stacked on a group of machine learning classifiers for multispectral images and a filter classifier for lidar point clouds. Multiple steps were required to obtain the final habitat map including: 1) Preprocessing; 2) Create Training Data; 3) Band Indices & Classification Methods; 4) Water Classification; 5) Agriculture Classification; 6) Lidar Classification; 7) Marsh, Algal Flat, Grass, Forest Classification; 8) Final Combination. DATA: To generate a habitat map for the West Matagorda Bay area, the following three sources were used. 1) 26 WV-2 tiles, with acquisition dates of 11/17/2012, 5/5/2013, and 12/16/2013, were needed to cover the study area. The imagery had 8 spectral bands with a 2-m spatial resolution, 2) Two lidar datasets were needed to cover the entire study area and were downloaded from the Texas Natural Resources Information System. Acquisition dates ranged from 1/4/2018 – 2/23/2018 and 1/24/2019 – 1/29/2019, and 3) NOAA’s 2016 Coastal Change Analysis Program (CCAP) Regional Land Cover dataset assisted in classifying the WV-2 imagery. C-CAP uses data from the National Land Cover Dataset (NLCD) to classify 25 land cover types with a spatial resolution of 30 x 30 m. PREPROCESSING: Each WV-2 image was corrected for terrain displacement and radiometric settings by the Polar Geospatial Center (PGC) at the University of Minnesota. Then, University of South Florida’s MATLAB code was used to 1) Radiometrically calibrate digital count data, 2) Atmospherically correct images by subtracting Rayleigh Path Radiance, and 3) Convert images to surface reflectance by accounting for Earth-Sun distance, solar zenith angle, and average spectral irradiance. CREATE TRAINING DATA: First, training data was created. ENVI software’s Feature Extraction was used to generate segments based on similar spectral signatures. Segments were then manually classified into three land covers: 1) bare soil, 2) shrub/tree/grass, and 3) marsh/algal flat. BAND INDICES AND CLASSIFICATION METHODS: Three band indices were used to classify land cover: 1) Structure Insensitive Pigment Index (SIPI), 2) Shadow Index (SI), and 3) Modified Soil Adjusted Vegetation Index (MSAVI). The training data classifications in combination with the three band indices (SIPI, SI, and MSAVI) were ran through five pixel-based classification methods (Random Forest, Support Vector Machine, Multilayer Perceptron, Maximum Likelihood, Gradient Boosting Machine) to classify each pixel into three land covers (bare ground, upland grass/forest, and marsh/algal flat). For each pixel, the results from the five pixel-based classification methods were combined (based on how consistently each land cover class was classified in each method) to obtain final, more accurate land cover designations. WATER CLASSIFICATION: WV-2 was used to create a Normalized Difference Water Index (NDWI) which resulted in a continuous water body (gulf and bay) plus numerous isolated water bodies (lakes). Due to presence of multiple water bodies, the Connected Components Labeling (CCL) algorithm was used to group water pixels together. Following the execution of the CCL algorithm, water was extracted from 2016 CCAP data and reduced by 60 m. Next, the CCL algorithm was used to group pixels together based on similar elevation values. Then, the CCL results from WV-2 water and CCAP water were intersected to check for 1) complete, 2) partial, and 3) no overlap between water bodies. AGRICULTURE CLASSIFICATION: First, “Cultivated Crops” were extracted from the CCAP data and shrank by 60 m. Then, various steps were taken to cleanup the CCAP agriculture classification. Following the data cleanup, training pixels were randomly selected using the WV-2 spectral indices (SIPI, SI, MSAVI) outputs. Using the training pixels, the random forest classifier was used to classify agriculture. Lastly, the WV-2 classification pixels were grouped together using the CCL algorithm and the WV-2 agriculture classification was overlapped with the CCAP “Cultivated Crops” classification. If the WV-2 agricultural classification block overlapped the CCAP “Cultivated Crops”, the WV-2 agricultural area was kept. If no overlap occurred, the WV-2 agriculture classification was removed. LIDAR CLASSIFICATION: First, holes were filled in the data. Then a digital elevation model (DEM), digital surface model (DSM), and normalized digital surface model (nDSM) were calculated. Next, all unclassified lidar points were used to create CCL. Based on the CCL values, buildings and vegetation were separated. Lastly, vegetation was separated into three classes based on their values: 1) low vegetation (CCL < 0.75 m), 2) medium vegetation (CCL 0.75 – 2 m), and 3) high vegetation (CCL > 2 m). MARSH, ALGAL FLAT, GRASS, FOREST CLASSIFICATION: After the initial classification of bare ground, grass/tree, and marsh/algal flat from the three WV-2 derived band indices and five pixel-based classification methods, additional steps were needed to separate 1) upland grass from trees, 2) marsh from algal flat, and 3) any other misclassified pixels. A complex decision tree incorporating the lidar low, medium, and high vegetation returns and the original land cover designations was developed to fix the aforementioned issues. FINAL COMBINATION: After all classifications were complete, the individual classification rasters were combined to get one raster for each WV-2 image. The order of raster combination/overlap was crucial to obtaining the most accurate classification and was overlapped as follows: 1) Marsh, Algal Flat, Upland Grass, and Upland Forest Classification, 2) Agriculture Classification, 3) Water Classification, and 4) Culvert, Bridge Deck, and Building Classification. Once the classifications were combined into one raster for each WV-2 image, salt-and-pepper artifacts that arise when conducting pixel-based classifications were eliminated. Specifically, small features <= 9 pixels (36 m2) were removed from water bodies. Lastly, after executing the above processing steps and getting final classifications for the 26 WV-2 images, the 26 final classification outputs were mosaicked together by giving preference to the newer imagery dates in overlap areas.

Error Analysis:

A classification accuracy assessment was conducted to assess the accuracy of the WV-2 classifications. A 50 m buffer was generated around the CMGL’s most recent ESI shoreline and validation points were created with at least 50 m between points using ArcGIS’ “Create Random Points” tool. Ultimately 493 randomly located points were created and were manually classified by referencing the WV-2 imagery and Pictometry’s oblique imagery.

Provenance and Historical References:

NOAA. (2016). C-CAP Regional Land Cover. Retrieved from https://coast.noaa.gov/digitalcoast/data/ccapregional.html Su, Lihong, Mukesh Subedee, Marissa Dotson, James Gibeaut, Brach Lupher, Anthony Reisinger, and Rhiannon Bezore. 2021a. 2-meter Topographic Lidar Digital Elevation Model (DEM) of the Lower Texas Coast. Distributed by: Gulf of Mexico Research Initiative Information and Data Cooperative (GRIIDC), Harte Research Institute, Texas A&M University–Corpus Christi. doi:10.7266/Z7WG9EGN. Retrieved from https://data.gulfresearchinitiative.org/data/HI.x833.000:0010 Su, Lihong, Mukesh Subedee, Marissa Dotson, James Gibeaut, Brach Lupher, Anthony Reisinger, and Rhiannon Bezore. 2021b. 2-meter Topographic Lidar Digital Elevation Model (DEM) of the Upper Texas Coast. Distributed by: Gulf of Mexico Research Initiative Information and Data Cooperative (GRIIDC), Harte Research Institute, Texas A&M University–Corpus Christi. doi:10.7266/2MYPTJ7Y. Retrieved from https://data.gulfresearchinitiative.org/data/HI.x833.000:0009 USGS. (2018). South Texas Lidar. Retrieved from https://data.tnris.org/collection?c=6131ecdd-aa26-433e-9a24-97ac1afda7de#6.68/27.576/-98.187 USGS. (2019). Matagorda Bay Lidar. Retrieved from https://data.tnris.org/collection?c=8774ed51-b633-4f03-85ca-94c311ee0a88#8.66/28.7509/-96.1562 The WV-2 images used are available from Apollo Mapping http://apollomapping.com/.

Individual Files: