Clustering of spectral images using Echo state networks

Abstract

In the present work we applied a recently developed procedure for multidimensional data clustering to processing of spectral satellite images. The core of our approach lays in projection of multidimensional image to a two dimensional one. The main aim is to discover points with similar characteristics. This was done by clustering of the resulting image. The processing technique exploits equilibrium states of a kind of recurrent neural network -Echo state network (ESN) -that are obtained after intrinsic plasticity (IP) tuning of the ESN using multidimensional data as inputs. The proposed in our previous work automated procedure for multidimensional data clustering is further refined and tested on the satellite image data. The obtained number and position of clusters of a multispectral image of a mountain region in Bulgaria is compared with the classification of the region landscape given by the Ministry of Regional Development and Public Works.

An experimental evaluation of echo state network for colour image segmentation

“…For each image and for each chosen triplet of the reservoir parameters, the experiment is repeated five times where the maximum standard deviation for all images is 0.0176 (corresponding to a mean F-score of 0.8033). This small value of the standard deviation proves that even random selection of a small number of neurons from the reservoir can provide adequate features for good quality image segmentation, unlike the work in [12], discussed in section II, where the selection of neurones from the reservoir requires an excessive computational load.…”

“…Koprinkova et al [12] have used the ESN as features extractor for clustering of multi-spectral satellite images. Firstly, the ESN reservoir parameters are improved using the intrinsic plasticity (IP) based adaptation.…”

“…One possible realization of RC is the echo state network (ESN) proposed by Herbert Jaeger [7], which consists of a large randomly generated untrained RNN followed by a readout layer that is trained using linear regression algorithms. Despite the simplicity of the ESN concept, it has been successfully applied in many tasks such as time series prediction [8], traffic load prediction [9], detection of weak signal embedded in chaotic background [10], image segmentation [11], spectral images clustering [12] and microscopic cellular image segmentation [13]. Another similar approach to ESN is the Liquid State Machine (LSM) which has been proposed by Maass et al [14].…”

Echo state network‐based feature extraction for efficient color image segmentation

“…Distinguished by the simplicity of its nodes and the ease of its training, the ESN has been used in many engineering applications. [9][10][11][12][13] Despite its simple architecture and ease of implementation, ESN configuration requires some practice and insight to obtain a good performance in many applications. 14 Several studies have been carried out to explore the ESN parameters and evaluate their performance in many engineering tasks.…”

“…14 Several studies have been carried out to explore the ESN parameters and evaluate their performance in many engineering tasks. [15][16][17] However, in spite of the application of ESN to image segmentation in previous works, 12,13 to the best of our knowledge, a thorough investigation of its applicability to color image segmentation is still lacking. In addition, most of the applications based on ESN process a temporal data.…”