Remote Sensing Image Classification with a Few Labeled Samples

Hyperspectral imaging (HIS) provides a strong base for classification in the remote sensing because it compromises a lot of spectrums as well as geospatial data. The HSI classification accuracy is influenced by the high dimensionality of the hyperspectral data, The main of the project is to extraction of non-redundant, uncorrelated and noise-free bands from the remote sensing images, to device a robust algorithm for classification of hyperspectral image using machine learning approaches. The method propose was proposed for the extraction of the representative bands from high-dimensional imagery in order to improve classification. To make the most HIS data, Factor Analysis is used in the article to lower the dimensions of HSI. Furthermore, the research’s classification block is a convolution neural network (CNN), which was applied to the system to get improved classification results performance with deeper layers. In CNN model, ReLU activation function is used for the last layers and SoftMax activation function for the middle layer. The overall accuracy for principal component analysis is 99.1 and for factor analysis is 99.6. Extensive tests on Indian Pines data sets show that the suggested strategy enhances HSI classification performance considerably.

I. Introduction

Remote sensing hyperspectral images (HSIs) are acquired at a significant number of discrete and continuous spectra wavebands, which typically span from 0.4m to 2.5m and encompass the electromagnetic spectrum from visible light to (near) infrared. This makes HSIs more useful in a variety of applications, such as mining, agricultural, geoscience, and surveillance systems. HSI also validates counterfeit items and papers in addition to assisting with food quality, pharmacological, defense, and skin assessment [1]. HSI offers superior discriminability in data processing for numerous humanistic facilities like discriminatory practices between vegetation classes, identification of the inclusion or exclusion of particular components in a soil or rock, and so on, with its spectral resolution in nm. Generally, HSI data includes a hypercube that includes 2D spatial features and image features in the 3d space. As a result, the overall information stored is XYF, where F is the number of picture bands in the image, X is the no. of rows, and Y is the no. of columns for each image band [2]. Because the photos are typically obtained by airplane or satellite sensors, some processing is required, like environmental, radiological, and geometrical correction [3].