Science Service System

Summary of Proposal MTH3368

TitleConvolutional Neural Networks for contextual denoising and classification of SAR images
Investigator Danilla, Carolyne - University of Twente, Geo-information and Earth Observation
Team MemberNo team members defined
SummaryThe main objective of the proposed research is to investigate a new method adopting a deep learning approach based on convolutional neural networks (CNN), which is capable of learning from the data the speckle reduction filters, extract spatial-contextual features and for classification of SAR images. For this, a new CNN classifier / algorithm that work with raw unfiltered images will be developed and implemented with a multi-temporal crop type classification. The results will be compared with those of existing alternative methods such as support vector machines that work with speckle filtered images, and handcrafted features such as the texture features extracted by computing statistical parameters of grey level co-occurrence matrix (GLCM). The study site is a predominantly agricultural area –Flevoland with regular crop field boundaries and about six or more different crop types in a typical growing season. Since contextual information is important in this study, resolution is key in a way that spatial dependence is more pronounced at higher resolution than lower resolution and this effect will be investigated as well as the impact of resolution for precision agriculture.The will also highlight the importance of SAR images in Agricultural monitoring which usually requires multi-temporal classification (at different stages of growth) yet optical images have unstable temporal coverage due effect of cloud cover. Therefore dual polarized TerraSAR-X images are required at different growth periods. The results of this research as well as conclusions and recommendations will be detailed in a thesis report to be submitted and defended for the award of master’s degree in Geoinformation science and Earth Observation at the University of Twente

Back to list of proposals

© DLR 2004-2016