Mangrove extraction from super-resolution images generated by deep learning models
Mangroves are an essential component of coastal ecosystems. Accurate and effective identification and extraction of mangrove areas from remote sensing imagery is crucial for monitoring changes in the nearshore ecological environment. High-resolution remote sensing imagery is often difficult or expen...
| Main Authors: | , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2024-02-01
|
| Series: | Ecological Indicators |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1470160X24001717 |
| _version_ | 1827341756281126912 |
|---|---|
| author | Yu Hong Xiang Que Zhe Wang Xiaogang Ma Hui Wang Sanaz Salati Jinfu Liu |
| author_facet | Yu Hong Xiang Que Zhe Wang Xiaogang Ma Hui Wang Sanaz Salati Jinfu Liu |
| author_sort | Yu Hong |
| collection | DOAJ |
| description | Mangroves are an essential component of coastal ecosystems. Accurate and effective identification and extraction of mangrove areas from remote sensing imagery is crucial for monitoring changes in the nearshore ecological environment. High-resolution remote sensing imagery is often difficult or expensive to obtain and usually lacks sufficient temporal coverage, so most monitoring of mangrove forests still relies on medium- or low-resolution imagery, resulting in inaccurate distribution of extracted mangrove areas. The super-resolution (SR) images generated by increasingly widely used deep learning (DL) models may be an alternative. To validate this, we evaluated the extraction of mangroves from SR images generated by four DL models (i.e., enhanced super-resolution generative adversarial networks (ESRGAN), Real-ESRGAN, vast-receptive-field pixel attention network (VapSR), and image restoration methods using swin transformer (SwinIR)). The models were trained on paired (high-resolution) Chinese GF-1 satellite images and (low-resolution) Landsat 8 datasets. The performance of model fitting, vegetation indices, and extracted mangrove areas was evaluated using three commonly used classifiers (i.e., support vector machines (SVM), random forest (RF), and gradient-boosted decision tree (GBDT)) in combination with ground-truth sampling points and public mangrove datasets. Results showed that: (1) the SR images generated by DL-based models can facilitate the extraction of mangrove features. (2) The quality evaluation metrics peak signal-to-noise ratio (PSNR) and structural similarity index measurements (SSIM) cannot be regarded as absolute criteria for SR images, especially when the SR images were used for mangrove feature extractions. (3) The SR image generated by the VapSR model was best suited for extracting mangrove forests and performed even better than the original high-resolution images. (4) Taking a public dataset as a benchmark, the mangrove areas extracted from the SR image generated by the VapSR model were closer to the benchmark than the original Landsat-8 and GF-1. Overall, the DL-based SR models can enhance mangrove extractions and potentially have widespread applications. |
| first_indexed | 2024-03-07T21:53:34Z |
| format | Article |
| id | doaj.art-ac429dd2194e44488a082b6fe147ccc9 |
| institution | Directory Open Access Journal |
| issn | 1470-160X |
| language | English |
| last_indexed | 2024-03-07T21:53:34Z |
| publishDate | 2024-02-01 |
| publisher | Elsevier |
| record_format | Article |
| series | Ecological Indicators |
| spelling | doaj.art-ac429dd2194e44488a082b6fe147ccc92024-02-25T04:35:18ZengElsevierEcological Indicators1470-160X2024-02-01159111714Mangrove extraction from super-resolution images generated by deep learning modelsYu Hong0Xiang Que1Zhe Wang2Xiaogang Ma3Hui Wang4Sanaz Salati5Jinfu Liu6College of Forestry, Fujian Agriculture and Forestry University, Fuzhou 350002, ChinaDepartment of Computer Science, University of Idaho, Moscow, ID 83844, USA; College of Computer and Information Sciences, Fujian Agriculture and Forestry University, Fuzhou 350002, China; Corresponding author at: Department of Computer Science, University of Idaho, Moscow, ID 83844, USA.Department of Computer Science, University of Idaho, Moscow, ID 83844, USADepartment of Computer Science, University of Idaho, Moscow, ID 83844, USADepartment of Geosciences, Mississippi State University, Starkville, MS 39762, USADepartment of Computer Science, University of Idaho, Moscow, ID 83844, USA; Lowell Institute for Mineral Resources, University of Arizona, Tucson, AZ 85721, USACollege of Forestry, Fujian Agriculture and Forestry University, Fuzhou 350002, ChinaMangroves are an essential component of coastal ecosystems. Accurate and effective identification and extraction of mangrove areas from remote sensing imagery is crucial for monitoring changes in the nearshore ecological environment. High-resolution remote sensing imagery is often difficult or expensive to obtain and usually lacks sufficient temporal coverage, so most monitoring of mangrove forests still relies on medium- or low-resolution imagery, resulting in inaccurate distribution of extracted mangrove areas. The super-resolution (SR) images generated by increasingly widely used deep learning (DL) models may be an alternative. To validate this, we evaluated the extraction of mangroves from SR images generated by four DL models (i.e., enhanced super-resolution generative adversarial networks (ESRGAN), Real-ESRGAN, vast-receptive-field pixel attention network (VapSR), and image restoration methods using swin transformer (SwinIR)). The models were trained on paired (high-resolution) Chinese GF-1 satellite images and (low-resolution) Landsat 8 datasets. The performance of model fitting, vegetation indices, and extracted mangrove areas was evaluated using three commonly used classifiers (i.e., support vector machines (SVM), random forest (RF), and gradient-boosted decision tree (GBDT)) in combination with ground-truth sampling points and public mangrove datasets. Results showed that: (1) the SR images generated by DL-based models can facilitate the extraction of mangrove features. (2) The quality evaluation metrics peak signal-to-noise ratio (PSNR) and structural similarity index measurements (SSIM) cannot be regarded as absolute criteria for SR images, especially when the SR images were used for mangrove feature extractions. (3) The SR image generated by the VapSR model was best suited for extracting mangrove forests and performed even better than the original high-resolution images. (4) Taking a public dataset as a benchmark, the mangrove areas extracted from the SR image generated by the VapSR model were closer to the benchmark than the original Landsat-8 and GF-1. Overall, the DL-based SR models can enhance mangrove extractions and potentially have widespread applications.http://www.sciencedirect.com/science/article/pii/S1470160X24001717Mangrove extractionSuper-resolution modelsDeep learningRemote sensing |
| spellingShingle | Yu Hong Xiang Que Zhe Wang Xiaogang Ma Hui Wang Sanaz Salati Jinfu Liu Mangrove extraction from super-resolution images generated by deep learning models Ecological Indicators Mangrove extraction Super-resolution models Deep learning Remote sensing |
| title | Mangrove extraction from super-resolution images generated by deep learning models |
| title_full | Mangrove extraction from super-resolution images generated by deep learning models |
| title_fullStr | Mangrove extraction from super-resolution images generated by deep learning models |
| title_full_unstemmed | Mangrove extraction from super-resolution images generated by deep learning models |
| title_short | Mangrove extraction from super-resolution images generated by deep learning models |
| title_sort | mangrove extraction from super resolution images generated by deep learning models |
| topic | Mangrove extraction Super-resolution models Deep learning Remote sensing |
| url | http://www.sciencedirect.com/science/article/pii/S1470160X24001717 |
| work_keys_str_mv | AT yuhong mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT xiangque mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT zhewang mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT xiaogangma mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT huiwang mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT sanazsalati mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels AT jinfuliu mangroveextractionfromsuperresolutionimagesgeneratedbydeeplearningmodels |