A feature sequence-based 3D convolutional method for wetland classification from multispectral images. Issue 9 (1st September 2020)

• Record Type: Journal Article
• Authors: Pan, Hong
• Is Part Of: Remote sensing letters. Volume 11:Issue 9(2020)
• Page Start: 837
• Page End: 846
• Publication Date: 2020-09-01
• Subjects: Remote sensing -- Periodicals; Remote sensing; Periodicals; 621.3678
• DOI: 10.1080/2150704X.2020.1772518
• Languages: English
• ISSNs: 2150-704X

Abstract

ABSTRACT: As an important part of the ecosystem, wetlands provide varies of ecological functions while they have also been increasingly threatened and degraded. Therefore, it is necessary to protect wetlands with effective monitoring measures such as classification. Considering the quick development of multispectral sensors, multispectral images are available to classify wetlands. The current method, however, requires large quantities of samples and can therefore only be applied in some categories. The gist of this study is to propose and evaluate a feature sequence-based three-dimensional (3D) convolutional method for wetland classification from multispectral images. First, 3D convolution kernels based on feature sequence are designed. Then, an optimal 3D convolutional neural network framework is constructed after fine-tuning, based on which the training and initial classification are executed successively. Next, a correction algorithm based on data similarity matching theory is introduced to correct the misclassifications. Finally, the performance of each step has been evaluated and classification results of the proposed method have been compared with other existing methods. Results have shown that the overall accuracies of the proposed method have exceeded 74% for the research categories and it performs better than other methods when there are complex wetlands combined of multiple contents.