Application of a Hybrid Unsupervised Method Based on K-Means and Neural Networks for Satellite Image Classification and Image Quality Enhancement (Case Study: Sabzevar City, 2021)

Introduction: Nowadays, satellite imagery enables land cover classification, which is one of the key approaches in analyzing environmental changes, managing natural resources, evaluating land-use changes, and monitoring natural hazards. These methods help urban decision-makers, land resource managers, and others to obtain accurate information on the current situation and environmental change trends. Among these, unsupervised classification methods are widely used due to their independence from labeled training data, especially in the analysis of large and complex datasets. However, many clustering algorithms such as K-means, although computationally fast and simple, fail to deliver sufficient accuracy when facing challenges such as spectral overlap between classes, image noise, and complex spatial structures. These algorithms often cannot effectively model nonlinear features and complex relationships among different bands. On the other hand, artificial neural networks, particularly with proper design, offer strong capabilities in learning complex patterns and hidden structures within the data.

Objective: The aim of this study is to design and implement a fast, structured, accurate, and time-efficient method for unsupervised classification of Landsat 8 satellite imagery, while also contributing to the reduction of redundant sensors on satellites. This method is based on multiple indices and the integration of the K-means algorithm with a simple artificial neural network, designed in such a way that it does not require training data. One of the main goals of this research is to utilize combined features, including spectral bands (1–7), thermal bands (10 and 11), and the panchromatic band (band 8), along with common land cover indices such as NDVI, NDBI, and NDWI. This combination helps distinguish difficult classes such as dry soil, bare mountains, white rooftops, and spectrally similar urban surfaces. Moreover, intra-urban and inter-urban roads were separated with high accuracy. Another objective of this research is to reduce the negative effects of shadow, noise, and spectral mixing of pixels in urban–mountainous areas, thereby reconstructing spatial structures with high precision.

Methodology: In this approach, the Landsat 8 image comprising 10 main bands (7 multispectral, 2 thermal, and 1 panchromatic) was first pre-processed. Then, a variety of indices such as NDVI (vegetation index), NDBI (built-up index), NDWI (water index), SAVI, and DBI were extracted. Additional features such as brightness, texture, and edges were also derived for better class separation. After data normalization, the K-means algorithm was applied for initial clustering, and its results were used as weak labels for the neural network. The designed neural network included a simple hidden layer, which refined the clusters and performed the final classification by receiving the extracted features as input. The thermal band was employed to distinguish warm and dry soil surfaces from shaded and cold areas. The panchromatic band, with its higher resolution, played a key role in the precise detection of intra-urban and inter-urban roads.

Results: The final classification results demonstrated that the proposed method achieved higher quality compared to other methods in distinguishing spectrally similar classes such as rooftops, bare mountains, bare soil, non-covered surfaces, and intra-urban streets versus inter-urban roads with clear boundaries. The noise in the output of the improved unsupervised classification was significantly reduced, and spatial cohesion of classes was preserved. Qualitative evaluation in the Sabzevar region showed that the improved unsupervised classification successfully identified vegetation cover, built-up areas, and barren lands with high spatial accuracy and detail. Moreover, intra-urban streets were effectively separated from inter-urban roads, and the effects of shadows and noise in mountainous regions were reduced.

Innovation: The main innovation of this research lies in the step-by-step, novel, and integrated framework that, without relying on training data, delivers accurate results solely through the combination of K-means and a simple neural network. The use of extracted indices and the design of artificial bands to enhance class separability, along with the extraction of specific features such as texture and image edges, significantly improved classification accuracy. Another major innovation is the precise separation of urban roads from dense built-up structures and the complete distinction of built-up areas from surrounding lands. A noteworthy point is that this method was fully implemented in MATLAB without the use of pre-built functions or library tools, making it a lightweight, native, and transferable approach applicable to other images and regions.