Identification of Urban Areas in Densely Forested Areas of Puerto Rico Using Image Segmentation

Pixel-based Clustering of Densely Vegetated Population Centers with an ISODATA algorithm

December 2023

Locating Populated Areas Through Thick Vegetation

Classifying urban scenes in densely forested areas can be challenging due to the small scale of high albedo surfaces that usually occur in urbanized environments. Being able to locate population centers in remote areas that are at high risk for natural disasters is vital to ensuring that the people living in those areas have access to humanitarian aid.

Aguas Buenas is a township of Puerto Rico located in the island's central mountain range. Despite being next to very dense population centers, the town's population is spread throughout densely vegetated areas that are remote enough to be cut off from most major roads during hurricanes due to flooding, mud slides, and other secondary effects of high severity weather. A Landsat 8-9 OLI/TIRS C2 L2 image with minimal cloud cover of the Eastern side of Puerto Rico was cropped to the extent of the township of Aguas Buenas. Image segmentation of this product was carried out on ArcGIS pro using the unsupervised, pixel-based clustering ISODATA algorithm. Classified imagery was ground truthed by creating an equalized random stratified sample of validation sites. Unsupervised classification shows great potential to be a fast and accurate way of locating populated areas even through dense vegetation, which could help identify and take into account these areas when planning emergency management plans.

Methods

A Landsat 8-9 OLI/TIRS C2 L2 image with minimal cloud cover of the Eastern side of Puerto Rico taken on May of 2023 was cropped to the extent of the township of Aguas Buenas and prepared for analysis with ArcGIS Pro using the R Studio package Terra. The image was classified using the pixel-based clustering ISODATA algorithm inside ArcGIS Pro and then the results of the unsupervised classification were reclassified to represent three meaningful features within the image– Urban, Dense Forest Cover and Thin Forest Cover. The algorithm used 15 classes and had a minimum class size of 60 with a sample interval of 1. Pixel clusters were then reclassified using the spatial analyst tool Reclassify in ArcGIS Pro, and meaningful features were assigned a number from 1-3 for Urban, Dense Forest Cover and Thin Forest Cover respectively. These classes were then filtered using a majority filter with a queen weight's scheme to remove the salt and pepper effect of pixel based clustering. Data was then validated by extracting an equalized random stratified sample of 100 validation sites from the reclassified layer, and a confusion matrix was generated using classified and ground truthed data extracted from the validation sites layer.

Landsat 8-9 OLI/TIRS C2 L2 of Eastern Puerto Rico and surrounding islands

Results

Satellite Imagery

Landsat 8-9 OLI/TIRS C2 L2 imagery of the eastern side of Puerto Rico cropped to the extent of the town of Aguas Buenas.

Final Product

Pixel-based ISODATA Image run through a queen weights scheme majority filter.

ISODATA Results

Satellite Imagery run through unsupervised, pixel-based ISODATA algorithm.

The algorithm had a total precision of 0.72, with individual precision scores of 0.87 (Urban), 0.90 (Dense Forest Vegetation), and 0.39 (Thin Forest Vegetation). Recall was 0.65 (Urban), 0.76 (Dense Forest Vegetation), and 0.81 (Thin Forest Vegetation), with a calculated Kappa coefficient of 0.59 overall.

abreclass_val_cmatrix

Land Use Classification

The algorithm had a hard time distinguishing between thin forest vegetation and urban areas, an issue that probably stemmed from the fact that most properties in more remote populated areas of Aguas Buenas have a lot of land that has been turned into what can be classified as "Thin Forest Vegetation." When ground truthing with the validation sites, it was evident that these areas of thin forest vegetation were actually inside urban areas, so while they weren't technically incorrectly classified, that's still something to keep in mind.

Because of this, the reclassification could also benefit from having a broader range of land use classes, as with the remoteness of most housing it's important to distinguish between suburban and fully urbanized areas, as that could help with the recall of the algorithm.

Additionally, the southwestern part of the image was dotted with cloud cover, which was improperly identified as an Urban area due to the similarities in albedo between concrete, the primary building material for housing in Puerto Rico, and cloud cover. Martinuzzi et al. (2007) have developed a simple and semi automated method to mask clouds and shadows in Landsat imagery that they used to create cloud-free composites of multitemporal images for Puerto Rico and its adjacent islands. Such techniques could be applied in the future to create an image for classification that would be entirely free of cloud cover.

Unsupervised classification and image segmentation shows great potential to be a fast and accurate way of locating populated areas even through dense vegetation, which could help identify and take into account these areas when planning emergency management plans. With the support of local municipalities, GIS is a powerful tool that can be used to help those even in the most remote places.

Sources

"Municipio de Aguas Buenas" pr.gov. Accessed November 30, 2023.

Martinuzzi, Sebasti n; Gould, William A.; Ramos Gonzalez, Olga M. 2007. Creating cloud-free Landsat ETM+ data sets in tropical landscapes: cloud and cloud-shadow removal. U.S. Department of Agriculture, Forest Service, International Institute of Tropical Forestry. Gen. Tech. Rep. IITF-32.

Page updated

Google Sites

Report abuse