Multi-Scenario Target Detection using Neural Networks on Hyperspectral Imagery

Deep learning techniques play the lead role in target detection in remote sensing when compared to conventional and machine learning techniques. Target detection techniques have caught attention in many applications, like military, science, medicine, and many more. A noval Convolution Neural Network (CNN) model with different layers designed, which helps to exploit study areas and to implement target detection. Finally, output threshold image states where each pixel is a target or otherwise. This study attempted the pixel pair technique by increasing the training pixel count to improve the model’s performance. The machine learning techniques like Support Vector Machine (SVM) & Logistic regression in python and conventional methods like Spectral Angle Mapper (SAM), Matched Filtering (MF), Adaptive Coherence Estimator (ACE), Constrained Energy Minimization (CEM), Normalized Euclidean Distance (NED), and Spectral Similarity Mapper (SSM) are carried out using target detection tools in the Environment for Visualising Images (ENVI) software. The different types of targets used to understand how they influence the target detection procedure by changes in size, shape, material, and a few other aspects. We identified three study areas with a unique experimental setup, i.e., large targets with different backgrounds (first), small targets with additional materials and shapes against the vegetation background (second), and small targets with a sand background (third). Two new ideas are attempted, i.e., to generate a model using an image captured from a ground-based instrument, then apply the same model to the airborne satellite image, and to apply the transfer learning technique by creating and using the model after changing the target locations.