Multi-Strategies inferred Convolution Neural Network for Land Cover Classification through Landsat 8 OLI Hyperspectral Images

Land Cover Classification of the hyperspectral images poses a significant importance to categorize the land classes and identifying changes using change detection techniques with respect to various seasons’ variations and climates variation. It leads to considerable challenges to current researchers as temporal variability of spectral, temporal and spatial characteristics of different reflectance bands of Landsat dataset. Land cover changes highly affects the current ecosystem, hence change information of land cover is significant for estimating the atmosphere. Computation of land cover changes in urban and rural region is highly complex due to its wide ranging degradation and high spatial and spectral changes. Further to enhance urban and rural land cover classification and mapping of the similar lands, a discriminative deep learning model entitled as Multiple Strategies inferred Convolution Neural Network for extraction of end member is been designed and experimented in this article. Initially, end members of the hyperspectral images are exploited on computation of pure spectral signatures along its fractional abundances map of pixel of the images. Extracted Endmember are projected to the principle component analysis technique to reduce the spectral feature as it impact the model with computational complexity. Principle component analysis is incorporated to generate reduced feature by retaining the most significant features. The reduced spectral and spatial features is transformed to deep learning model considered as multiple strategies inferred convolution neural network model. Convolution layer and max pooling layer is employed for feature processing and fully connected layer is employed to process for identification and mapping of the land cover along computing the change detection constraints and spectral indices (end members). Changes detection identification on the spectral indices classifies the land cover regions and computation of the spectral and spatial value with respect to N finder algorithm provides the pixel purity index. N finder technique considered as a change detection method. Experimental investigation of the proposed deep learning architecture is implemented using Landsat 8 OLI dataset and it performance is evaluated using generated spectral indices along the conventional approaches. Present architecture produces accuracy above 99% on comparing with conventional classification techniques.