Deep Learning with Synthetic Hyperspectral Images for Improved Soil Detection in Multispectral Imagery

Remote sensing images play a critical role in natural disaster prevention and damage estimation. With the fast advances in hyperspectral sensors, the hyperspectral imaging also known as imaging spectrometry is becoming a relatively recent trend. Hyperspectral images are expected to be more effective in object detection, environmental monitoring, urban growth monitoring, and damage assessment. In this paper, we present a 4 layers deep convolutional neural network (CNN) model for soil detection by using the combination of 80 synthetic hyperspectral bands and its original 8 multispectral bands which are collected by the WorldView-2 satellite. Our experiment result shows that by using the combined 80 synthetic hyperspectral bands and the original 8 multispectral bands, the area under the curve (AUC) scores of our CNN model for soil detection on the three testing images has been improved by 7.42% in average from 76.26% to 83.48%, as compared to the result by using the 8 multispectral bands alone. We also applied the CNN model onto a set of high-resolution data which is created by pan-sharpening the original multispectral bands and its synthetic hyperspectral bands, which quadrupled the spatial resolution of the combined synthetic hyperspectral bands. With the increased spatial resolution of the combined synthetic hyperspectral bands, the average AUC scores of our CNN model was furtherly improved by 10.02% from 81.44% to 91.47%. This significant improvement indicates that by using the pan-sharpened synthetic hyperspectral bands, the performance of CNN model for soil detection has been greatly improved, the synthetic hyperspectral bands with increased spatial resolution is a great alternative in improving the performance of object detection and classification in remote sensing applications.

I. Introduction

Remote sensing images play a critical role in natural disaster prevention and damage estimation, such as landslide detection and flood damage estimation. Multispectral images, which usually refers to images with 3 to 10 bands, have been widely and regularly used in the remote sensing area since the 1970s. With the fast advances of hyperspectral sensors, the hyperspectral imagery also known as imaging spectrometry, which contains many more bands with much narrower bandwidth (10-20 nm) than multispectral bands are expected to be more effective in target identification [1][2], land cover classification [3], anomalous materials and objects detection [4][5]. NASA is planning a mission that will have a hyperspectral imager [6][7] covering the whole Earth for vegetation monitoring.