| |
|
|
|
Flood Mapping Using Remote Sensing Data and Deep Learning-Based Segmentation Techniques |
|
|
|
PP: 39-49 |
|
|
|
doi:10.18576/isl/130503
|
|
|
|
Author(s) |
|
|
|
Ahmed Noureldeen,
Ahmed S. Ahmed,
Mohamed Z. Ghozl,
Amjad Feras,
Mennatallah Ahmed,
Naryman A. Ismail,
Sara Salah,
Shahd Ahmed,
Rawan Mohamed,
Ahmed Sallam,
|
|
|
|
Abstract |
|
|
| In this paper, we present an approach for classifying pixels as watery or non-watery in multispectral images for segmentation purposes. Our method begins with cloud detection using both k-means clustering and OpenCV thresholding techniques, although these factors had a minimal impact on our target outcomes. We implemented and evaluated various segmentation architectures, including U-Net, PAN, MA-NET, PSP-Net, DeepLabv3, and DeepLabv3+, achieving segmentation accuracy scores ranging from 74% to 94%. Among these, DeepLabv3 and DeepLabv3+ proved to be the most effective, each attaining a segmentation accuracy of 94%. The output from these models produced maps where watery pixels were designated at a value of 1, and non-watery pixels were designated a value of 0. |
|
|
|
|
 |
|
