ImageClassification using Deep Learning

IMAGECLASSIFICATION

Imageclassification refers to the task of extracting information classes from a multibandraster image. The resulting raster from image classification can be used tocreate thematic maps. Depending on the interaction between the analyst and thecomputer during classification, there are two types of classification:supervised and unsupervised.

#61557; Supervised vs Unsupervised classification

#61557; Supervised requires the analyst to identifyknown areas

#61557; Unsupervised determines a set number ofcategories based on a computer algorithm

#61557; Hybrid classifiers are a mix of the two

DEEP LEARNINGis a fascinating field. Artificial neural networks have been around for a longtime, but something special has happened in recent years. The mixture of newfaster hardware, new techniques and highly optimized open source librariesallow very large networks to be created with frightening ease.

In my thesis, Iwill chose the best of breed Python deep learning library called Keras thatabstracted away all of the complexity. The thesis will use many of the top deeplearning platforms and libraries, choose what I think is the best-of-breedplatform for getting started and quickly developing powerful and evenstate-of-the-art deep learning models in the Keras deep learning library forPython.

ConvolutionalNeural Networks represent a powerful artificial neural network technique. Thesenetworks preserve the spatial structure of the problem and were developed forimage classification. Convolution Neural Networks (CNNs) in essence are neuralnetworks that employ the convolution operation (instead of a fully connectedlayer) as one of its layers. CNNs are an incredibly successful technology thathas been applied to problems wherein the input data on which predictions are tobe made has a known grid-like topology. They are popular because people areachieving state-of-the-art results on di#64259;cult computer vision and naturallanguage processing tasks. Convolutional neural networks typically consist ofan input layer, a number of hidden layers, followed by a softmax classificationlayer. The input layer, and each of the hidden layers, is represented by athree-dimensional array with size, say, M × N × N. The second and thirddimensions are spatial. The first dimension is simply a list of featuresavailable in each spatial location. For example, with RGB color images N × N isthe image size and M = 3 is the number of color channels.