Efficient Classification and Soil Quantification Using Hyperspectral Data

Abstract

Hyperspectral data accommodates a large amount of information generally in the range of 400 - 2500 nm in the electromagnetic regions. It contains the reflectance values that are captured from the measuring device. This information is high dimensional in nature, which can be utilised in several real term applications. Classification of the hyperspectral image is a prevalent domain. Most of the existing classification techniques are not able to excerpt the deep features adjacent to the hyperspectral image as it comprises of a significant number of bands. So, we proposed a deep learning-based approach to extract the deep features efficiently. This approach is a hybrid of Locality Preserving Projection, Stacked Autoencoders, and Logistic Regression. Locality Preserving Projection is a dimensionality reduction procedure that is employed to decrease the hyperspectral input. The reduced input contains the preserved local information of the input hyperspectral image. Afterward, the reduced hyperspectral image is passed to the Stacked Autoencoders for taking out the essential approximations of the input that is regarded as deep features which are then passed to the Logistic Regression for efficient construction of the prediction model. Standard hyperspectral image dataset is utilised for validating the developed model results, and it is compared with 21 machine learning models to prove the efficacy of the proposed model. Most of the existing frameworks for classification are based on spectral-spatial information. The hyperspectral image is termed as high dimensional image because it comprises of many values corresponding to the bands and adding more spatial information to already a large dataset leads to more complex dimension set. So, we proposed a pre-processing framework by using spectral values set of the hyperspectral image. The spectral features are reduced with the two most popular dimensionality reduction algorithms named- Locality Preserving Projection and Principal Component Analysis. The feature set obtained from these algorithms are integrated to design a more efficient feature set called ’hybrid features’ which contains the local and global characteristics corresponding to the original hyperspectral image. The hybrid feature set has fewer dimensions than the original dataset due to which it takes less time for supervised training of the learning models. In addition to this, linear interpolation is incorporated in the framework to approximate the bands which are lost during the atmospheric correction process of the hyperspectral image which supports in procuring the better results in classification rate as well as training time in comparison with the spectral-spatial and spectral iii frameworks. The progress achieved in deep learning changed the conventional methods of analysing the information in every domain. It is rarely used in the field of soil science. A Principal Component Analysis and Long Short-Term Memory Networks based framework is proposed to perform the effective quantity estimation of soil characteristics from the spectral library. Firstly, the dimensions of the hyperspectral dataset are reduced with the help of Principal Component Analysis after pre-processing of the dataset. Then the output of Principal Component Analysis is passed to Long Short-Term Memory Networks for efficient learning. Long Short-Term Memory Networks perform better on sequential data as it can remember long term and short-term dependencies in the input. In the end, it is compared with other commonly employed models to show the effectiveness of the proposed model. We further improved the performance of the Principal Component Analysis and Long Short-Term Memory Networks framework by using the hybrid framework. Previously, we are only using the Principal Component Analysis features which denote the global approximation of the original dataset. So, local features obtained from the Locality Preserving Projection are combined with the Principal Components to form the hybrid features which are passed to Long Short-Term Memory Networks. It generated excellent results in comparison to the Principal Component Analysis and Long Short-Term Memory Networks framework due to efficient learning via hybrid features