Spectral discrimination using neural network for target detection

Remote sensing has enabled mapping, monitoring and management of many resources like water, agriculture; forestry etc. Hyperspectral imaging is a powerful tool in the field of remote sensing and has been used for many military applications like detection of landmines, target detection and discrimination about target and decoys etc. The objective of target detection algorithms is to analyze the image data and detect the targets of interest automatically or with very less human intervention. Major issues in target detection applications are spectral variability, noise, small size of target, complex backgrounds etc. Many of the detection algorithms do not work for difficult targets like small or dim targets, camouflages targets. These issues may results in false alarms. Thus in many military applications the target / background discrimination is an important issue, hence analyzing target's behavior in realistic environments is crucial for accurate interpretation of hyperspectral imagery. Using standard libraries for studying targets's spectral behavior suffers from a limitation that targets are measured in different environment than application. This study uses data measured at same time and location as the HSI image. Aim is to analyze spectrums of potential target materials in a way that each target can be spectrally recognized (or distinguished) from a mixture of spectral data. Also for the purpose of spectral discrimination no standard metric was found in research papers. Hence this study is an attempt to distinguish them using neural network. Results obtained points to need for proper band selection for target detection and forms the basis for future research