Detection of engineered surfaces using deep learning approach in AVIRIS-NG hyperspectral data

Hyperspectral remote sensing is opening new avenues for multitude of urban applications. This paper extends target detection method for extraction of engineered surfaces or urban targets, particularly roads and roofs. The study involves application of deep learning using AVIRIS-NG data. In pre-processing, generating ground reference image using Vertex Component Analysis (VCA) is done instead of using it for spectral unmixing of mixed pixels. Principal Component analysis (PCA) is carried out at a scale of 30,40 and 50 components for dimensionality reduction followed by implementation of Convolution Neural Network (CNN) for three window sizes (5,7 and 9). This deep learning measure is effective for high prediction and the results appear significantly higher in comparison to the literature. The time complexity increases with increase in PCA components and window size, making a compromise with accuracy. The study analyses least explored subset of AVIRIS-NG hyperspectral data of Udaipur region (India) to assist urbanisation.

Automated summary

This study extends the target detection method for urban applications using deep learning with AVIRIS-NG data, showing significantly higher results compared to existing literature. However, as the number of PCA components and window size increase, the time complexity rises, leading to a compromise with accuracy.