Spatial-Spectral Feature for Extraction Technique for Hyperspectral Crop Classification

V G Vani,Thippeswamy K

doi:10.17485/ijst/v15i2.1810

Abstract

Objectives: To present an extraction technique for the classification of the hyperspectral crop using the spatial-spectral feature. Methods: This paper presents a spatial-spectral feature extraction method employing the Image fusion technique and intrinsic feature extraction and a model for Improved Decision Boundary (IDB) using Support Vector Machine (SVM). Findings: The experiments have been conducted by using the Indian pines dataset which was extracted using the AVIRIS sensor. The dataset comprises of 16 distinctive classes such as corn, wheat, oats etc, which have used for evaluation of our model. Before the evaluation of the dataset the model has been trained using different training datasets in order to increase the accuracy and reduce misclassification. Moreover, the Spatial-Spectral Feature (SSF) model aided in distinguishing between crop intrinsic features and shadow element under dynamic environment condition. Our model attained 99.54%, 99.4%, 99.25% and 9.8 sec for OA accuracy, AA accuracy, Kappa and Time respectively. Furthermore, the overall accuracy of the model for the Support Vector Machine-3-dimensional discrete wavelet transform (SVM- 3DDWT), Convolutional Neural Network and Spatial-Spectral Feature Extraction Technique showed result of 94.28%, 96.12% and 99.4% respectively. Other existing models showed a low accuracy for the same dataset. Further, for addressing class imbalance issues this work modelled an improved decision boundary model for SVM by considering soft-margin rather than hard margin. The SSF-IDBSVM model achieves much better accuracies with less misclassification in comparison with recent deep learning-based HSI classification model. Novelty: Recently, several feature extraction and deep learning-based crop classification model have been modelled. However, existing crop classification fails to distinguish crop intrinsic feature concerning shadow feature; further, consider hard decision boundary; as a result, high misclassification is induced for smaller class size. Hence, this study provides an extraction feature which provides the classification of the crop in less time with higher classification and less misclassification. Keywords: Artificial Intelligence; Datamining; Crop Classification; Feature Extraction; Feature Selection; Hyper Spectral Information; Machin Learning Technique

Highlights

The crop classification model is one of the principal components of agriculture crop checking by using hyperspectral imaging acquired through satellites
Performance evaluation of the proposed Spatial-Spectral Feature (SSF)-IDBSVM model is carried out, performance evaluation is carried out by comparing the proposed HSI-based crop classification model with the existing crop classification model[23]; the further model is evaluated on the data which is gathered from the AVIRIS sensor[31]
The evaluation is carried out on basis of average accuracy and overall accuracy [17]; overall accuracy shows the absolute classification over the complete test feature an average accuracy shows the individual class average; in the case of all the metrics higher value indicates the superiority of classification approach

Summary

Introduction

The crop classification model is one of the principal components of agriculture crop checking by using hyperspectral imaging acquired through satellites. Harvest planning through hyperspectral imaging arrangement help in settling on the different dynamic in farming climate, for example, yield determining, crop region evaluation, etc. Precision crop planning is vital and sways crop recognizable proof applications. Any obstacles that exist in crop identification utilizing hyperspectral imaging should be met [3]. The conceivable high measurement size of hyperspectral imaging information is as yet an open issue. Hyperspectral imaging information takes after high likenesses of surfaces, spectral signatures marks, and shapes among various yields. The existence of mixed pixels in hyperspectral imaging altogether sways the correctness of the existing hyperspectral imaging crop arrangement model

Objectives

Results

Discussion

Conclusion