Land Use and Land Cover Mapping Using RapidEye Imagery Based on a Novel Band Attention Deep Learning Method in the Three Gorges Reservoir Area
Open Access
- 23 March 2021
- journal article
- research article
- Published by MDPI AG in Remote Sensing
- Vol. 13 (6) , 1225
Abstract
Land use/land cover (LULC) change has been recognized as one of the most important indicators to study ecological and environmental changes. Remote sensing provides an effective way to map and monitor LULC change in real time and for large areas. However, with the increasing spatial resolution of remote sensing imagery, traditional classification approaches cannot fully represent the spectral and spatial information from objects and thus have limitations in classification results, such as the “salt and pepper” effect. Nowadays, the deep semantic segmentation methods have shown great potential to solve this challenge. In this study, we developed an adaptive band attention (BA) deep learning model based on U-Net to classify the LULC in the Three Gorges Reservoir Area (TGRA) combining RapidEye imagery and topographic information. The BA module adaptively weighted input bands in convolution layers to address the different importance of the bands. By comparing the performance of our model with two typical traditional pixel-based methods including classification and regression tree (CART) and random forest (RF), we found a higher overall accuracy (OA) and a higher Intersection over Union (IoU) for all classification categories using our model. The OA and mean IoU of our model were 0.77 and 0.60, respectively, with the BA module and were 0.75 and 0.58, respectively, without the BA module. The OA and mean IoU of CART and RF were both below 0.51 and 0.30, respectively, although RF slightly outperformed CART. Our model also showed a reasonable classification accuracy in independent areas well outside the training area, which indicates the strong model generalizability in the spatial domain. This study demonstrates the novelty of our proposed model for large-scale LULC mapping using high-resolution remote sensing data, which well overcomes the limitations of traditional classification approaches and suggests the consideration of band weighting in convolution layers.Keywords
Funding Information
- the National Key R&D Program of China (2016YFA0600301)
This publication has 45 references indexed in Scilit:
- Patterns and driving forces of cropland changes in the Three Gorges Area, ChinaRegional Environmental Change, 2012
- Object based image analysis for remote sensingISPRS Journal of Photogrammetry and Remote Sensing, 2010
- Assessing the influence of environmental impact assessments on science and policy: An analysis of the Three Gorges ProjectJournal of Environmental Management, 2009
- Optimization in multi‐scale segmentation of high‐resolution satellite images for artificial feature recognitionInternational Journal of Remote Sensing, 2007
- The Three Gorges Dam: an ecological perspectiveFrontiers in Ecology and the Environment, 2004
- A multi-scale segmentation/object relationship modelling methodology for landscape analysisEcological Modelling, 2003
- Decision tree classification of land cover from remotely sensed dataRemote Sensing of Environment, 1997
- NDWI—A normalized difference water index for remote sensing of vegetation liquid water from spaceRemote Sensing of Environment, 1996
- Neural Network Approaches Versus Statistical Methods In Classification Of Multisource Remote Sensing DataIEEE Transactions on Geoscience and Remote Sensing, 1990
- A soil-adjusted vegetation index (SAVI)Remote Sensing of Environment, 1988