IMPROVING TEXTURE RECOGNITION ACCURACY IN SATELLITE IMAGES USING DEEP LEARNING TECHNIQUES

Abstract

This paper addresses the issue of recognizing texture in satellite imagery by utilizing a hybrid approach that combines both traditional texture analysis methods and deep convolutional neural networks (CNNs). The overall framework combines CNN architectures with features derived from the Gray-Level Co-occurrence Matrix (GLCM) and wavelet transform coefficients to compensate for weaknesses found in conventional techniques in capturing complex spatial patterns. In the case of satellite imagery, the need for data augmentation and preprocessing were both identified, and they used data management strategies that addressed falsehoods of information (data dependence) and noise (data stabilization). The model was validated on the EuroSAT benchmark dataset and achieved a 94.7% classification accuracy–an increase of 12.4% of GLCM approaches, and 5.2% of the pure CNN approaches. Of particular note, this research demonstrated an improvement in clearly identifying visual similar classes of land cover, such as urban areas and agricultural landscapes, with accuracy levels above 93% for each class. Based on the work, the author proposes an optimized methodology for remote sensing applications in feature fusion, confirms the application of multi-scale texture representation in remote sensing, and presents real-world implications for researchers that would like to implement hybrid models on a budget. These results have implications in the field and domain of precision agriculture, urban planning, and environmental monitoring [15][16]