Optimal spectral band detection for wet farmland localization in sattelite images

Abstract

In this work, we compared the importance of spectral bands made by satellite detection of underground waters on the agricultural land. Precise estimation of the area affected by floods is of great importance for yield prediction and farmer's subsidies given by government agencies. As input data for our research we have used images generated by WorlView-2 satellite. The most important properties of this satellite are very high spatial resolution of 1.84m for multispectral images and four new spectral bands: coastal-blue, red-edge, yellow and near-infrared 2. High resolution of satellite is substantial for us, because our fields of interest are small parcels in Northern Serbia. For optimal spectral band detection for wet farmland we used Support Vector Machine algorithm with Gauss kernel functions. The results presented show that very good performance in wet farmland detection can be achieved with less than all 8 channels with proper selection of the most informative channels.