Geospatial Object Detection in Hyperspectral Imagery Using Spectral-Spatial Networks

By recording a wide range of spectral bands across the electromagnetic spectrum, hyperspectral imagery (HSI) delivers extensive information. Due to the complicated interaction between spectral and spatial properties, extracting meaningful geographical information from HSI is difficult. This article uses Spectral-Spatial Networks to detect geographical objects in hyperspectral data. The proposed deep learning system blends spectral and spatial information to improve object discrimination in complicated scenarios. SSN uses convolutional neural networks (CNNs) and spectral attention mechanisms to adaptively weigh spectral bands and capture spatial correlations. A large dataset of annotated hyperspectral photography with ground truth object identification trains the SSN. The collection includes different lighting, topography, and item scales. The object detection task is fine-tuned after pre-training on a large HSI dataset. Experimental results on benchmark hyperspectral datasets show that the SSN outperforms state-of-the-art approaches. Precision, recall, and F1-score for numerous object classes reveal that the suggested technique can distinguish objects in complicated and congested situations. The suggested SSN framework is also resilient to illumination, climatic conditions, and sensor properties, making it suitable for real-world applications. The SSN architecture's scalability is further examined, showing its suitability for heterogeneous HSI datasets with different spectral resolutions.