Sub-pixel mapping with neural networks : real-world spatial configurations learned from artificial shapes

摘要: Koen C. Mertens, Lieven P.C. Verbeke and Robert R. De WulfLaboratory of Forest Management Spatial Information TechniquesGhent University9000 Gent, Belgiumtel +3292646091 - fax +3292646042koen.mertens@ugent.be, lieven.verbeke@ugent.be, robert.dewulf@ugent.beKEY WORDS: Neural, Networks, Classification, Algorithms, MultiresolutionABSTRACT:Remotely sensed images contain pure mixed pixels. A sub-pixel classification defines the membership degree ofpixels for different land cover classes. Sub-pixel mapping uses this information to obtain an image with a finerresolution. Pixels are divided into sub-pixels, representing class fractions. neural network is trained tolearn appropriate locations sub-pixels belonging classes inside pixel. Algorithm developmentand training was performed on synthetic imagery. Testing degraded real imagery yielded acceptableaccuracy measures. There ample room improvement technique.1 INTRODUCTION1.1 MappingRemotely usually both mi-xed Crisp assigns pixels theclass highest proportion coverage or probabil-ity. Loss inevitable during process.Soft techniques were introduced compen-sate loss: they assign pixel fractions landcover corresponding represented area insidea soft yields number fractionimages equal How-ever, assignment these renders no informa-tion about location pixel.Atkinson (1997) stated that it possible frac-tions spatially so called ’sub-pixels’. Sub-pixels finer resolution representation parent can be either mixed. In work sub-pixelswill assumed pure, i.e. one only.Sub-pixel seen as technique usesthe present in attain ahigher representation. The aim accord from parts ofthe pixel: sub-pixels. Many alternative map-ping methods have been suggested using sharpening im-ages (Foody, 1998), knowledge based analysis techniques(Schneider, 1993), Hopfield (Tatem et al.,2001a, Tatem al., 2002), linear (Verhoeye Wulf,2002) nonlinear optimisation (Gross andSchott, Bayesian approach (Gavin Jennison,1997), segmentation (Aplin Atkinson, 2001), de-convolution filters (Ruiz Lopez, genetic al-gorithms (Mertens 2003). study presented heredescribes where artificial networkestimates positions thepixels.1.2 Neural NetworksThis introduces feed-forward back-propagation neu-ral purposes. This networkarchitecture commonly used wide range appli-cations. (soft) well knownremote sensing application networks (Moodyet 1996, Atkinson 1997, Tatnall,1997, Foody, 1999). howeveris clearly distinct mapping, beused input mapping. istrained learn most examples.The building blocks artifi-cial neurons. These neurons, like biological havemultiple inputs single output, though out-put may connected many other To computethe output neuron, each multiplied by weightfactor. sum weighted neu-ron activation. neuron calculated func-tion Very often, neurons within networkare grouped three more layers: layer, oneor intermediate hidden layers, layer.Each layer (except layer) then fully con-nected previous layer. inthis experiment restricted networks. Inthis type networks, feedback connections not pos-sible. common procedure find net-work weights standard algorithm.In algorithm, pattern first propa-gated through feed-forwardphase. Afterwards, difference between calculatedand desired back-propagated outputneurons network, thereby adjustingthe opposite direction deriva-tive error respect individualnetwork weight. By repeating multiple times eachpattern set, taught mapthe correct outputs. For comprehensive dis-cussion see Haykin (1999).THE INTERNATIONAL ARCHIVES OF THE PHOTOGRAMMETRY, REMOTE SENSING AND SPATIAL INFORMATION SCIENCES (CD-ROM), VOL XXXIV-7/W9, Regensburg, Germany, 27-29 June 2003