Dense Convolution Neural Network for Complex land Cover Classification and change detection using Landsat OLI 8 hyperspectral Image

Hyperspectral Image is employed to observe the earth surface with respect to spectral reflectance value on the electromagnetic spectral ranges of electromagnetic spectrum initiating on the visible regions to infrared region. It identifies the spectral feature space and classifies it with respect to minerals and land covers. Hyperspectral images have been proposed using existing unsupervised and supervised classifier towards classification. Multiple complications have been exploited on the hyperspectral images as it contains the Hughes phenomenon. In order to alleviate those problems, dense convolution neural network methodology is architected in this research. It capable of classifying the spectral images with reduced processing time. Initially pre-processing of the hyperspectral image is carried out to reduce the noise and improve the contrast of the hyperspectral images. preprocessed image is projected to linear discriminant analysis to extract the spectral end members and spatial endmembers of the image. Those extracted spectral and spatial endmembers is employed to Dense Connected Convolution Neural Network as it is highly capable in land cover classification and discrimination. Deep learning model containing the convolution layer and max pooling layer for feature map generation and fully connected layer is highly capable in generating the spectral indices to classify the land cover and detecting the land cover changes on the spatial features. Softmax layer employs the spectral index for classification and change detection. Cross entropy is employed to reduce the loss of the model. Further hyperparameter tuning is carried out using gradient decent to generate the high discriminant classes of the land cover. The experimental outcomes of the current model are verified on using hyperspectral image gathered from Landsat 8 OLI dataset. Performance of the current approach is evaluated with state of art approaches on measures like precision, recall and f measure on cross fold validation. On analysis of its outcomes, it is confirming that current architecture exhibiting increased performance in classification accuracy of 99.43% on compared with conventional approaches.