|
Improving the accuracy of classification of multispectral Images using an anisotropic diffusion neural network algorithm
|
Parviz Zeaieanfirouzabadi *   , Mohammdad Tavakkoli Sabour , Behnaz Torkamani Asl |
|
|
|
Abstract: (21 Views) |
Classification of remote sensing images for the interpretation and preparation of thematic maps is one of the most important topics in the science of remote sensing. Given the importance of this process in extracting information from multispectral satellite images, the present study aims to improve the accuracy of satellite image classification by utilizing the Anisotropic Diffusion Neural Network (ADNN) algorithm along with texture information obtained from wavelet transformation. This method modifies the pixel values of the input image with a sequential weighted average based on neighboring pixels in the input bands. In other words, this algorithm acts as a spatially and temporally variable low-pass filter. The mentioned algorithm transfers the input multispectral image to five levels of scale/resolution. This algorithm can simultaneously process the spectral information of the image and the texture details obtained from the wavelet transformation in a multi-scale representation. In this study, Landsat 8 Level 2 and Sentinel 2 images from the Miandoab area in West Azerbaijan province were used. After extracting different levels of images through the ADNN algorithm, the classification of the resulting images into four classes (soil, water, vegetation, and residential area) was carried out using the Support Vector Machine (SVM) algorithm. Additionally, to evaluate the performance of the images obtained from the ADNN algorithm on the original image, classification using the Artificial Neural Network (ANN) method was also conducted. Finally, by selecting the best level, unsupervised classification using FCM (Fuzzy C-means clustering) was applied to the selected image.
The results of this research indicate that the highest Kappa classification in Landsat 8 image belongs to the Level 2 image of the ADFN algorithm with an accuracy of 86%, which shows a higher accuracy compared to the classification of the Level 2 image with an Artificial Neural Network (ANN) with a Kappa of 0.83. Additionally, this algorithm also showed similar performance in the Sentinel 2 sensor image with a Kappa of 0.83 at Level 2. It should be noted that the use of this algorithm in remote sensing image classification has not been significantly addressed in internal and external articles. For instance, in previous studies, only wavelet transformation and neural network based on anisotropic diffusion were used for change detection. Also, the Superpixel Segmentation algorithm based on anisotropic diffusion (ADS) has been introduced to improve the boundary accuracy of superpixels and correct boundary deviations in complex remote sensing images. So far, the accuracy of this algorithm has not been compared in similar studies. This confirms the innovation of this research in applying the Anisotropic Diffusion Neural Network (ADNN) algorithm to improve the classification accuracy of satellite images. |
Article number: 6 |
| Keywords: Unsupervised classification, Anisotropic diffusion neural network, Multispectral Images, Multi-Level classification, Remote Sensning |
|
|
Full-Text [PDF 2395 kb]
(10 Downloads)
|
Type of Study: Research |
Subject:
Photo&RS
Received: 2022/08/27 | Accepted: 2026/03/16
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
