Effective Hyperspectral Image Classification Using Learning Models

Recently, machine learning has produced appreciable performance results on various visual computing related studies, including the classification of common hyperspectral images. This study aims to compare the results of different machine learning models for the classification of a hyperspectral image dataset. The hyperspectral data captured from AVIRIS sensor covering scene over the Indian Pines test site in North-western Indiana and consists 224 spectral reflectance bands. The ground truth has sixteen classes including vegetation crops, built structures, etc. Accuracy assessments and confusion matrices were used to evaluate classification performance. The study includes classification results of mainly three learning models including dimensionally reduced data via PCA for SVM classification, CNN and k-NN. The overall accuracy in PCA-SVM results was 72.38%, CNN was 85% and k-NN was 66.21% concluding the better efficiency of CNN classification for the hyperspectral dataset. Hence CNN classification technique succeeded in the hyperspectral image classification.