Band Selection Technique for Crop Classification Using Hyperspectral Data

Hyperspectral datasets are widely used in many applications of remote sensing in the fields of agriculture, forestry, weather, urban planning, water study, biodiversity and surface changes. Our focus is agriculture with different perspectives such as classifications of crops, identification of different crop growth stages, fallow land, etc. However, the major challenge in working with hyperspectral data is the large number of narrow bands to deal with while having lack of fully labelled ground-truth data sometimes. In this paper, we propose a new band selection algorithm using the statistical parameter, spectral information divergence (SID), from the hyperspectral dataset. A hypothesis of the proposed band selection algorithm is if two most similar crops can be discriminated based on the lowest SID value, then other crops can also be discriminated. Subsequently, we use three classifiers: support vector machine (SVM), K-nearest neighbours and artificial neural network (ANN) and present the performance of the algorithm using simulations with the two performance parameters, i.e. overall accuracy (OA) and kappa coefficient. We have applied the algorithm on the three datasets: AVIRIS-NG dataset of Anand district in the state of Gujarat of India, Indian Pines dataset of northern-western Indiana state of the USA and Salinas dataset of Salinas Valley of California state of the USA. We have obtained the OA of 97.55% with the selection of 20 bands using SVM in the case of the AVIRIS-NG dataset. If we increase the selection to 40 bands, there is a nominal improvement in OA, i.e. 98.40%. We have also presented the performance of some prevailing algorithms working on band correlation analysis on the same datasets. We conclude that the proposed algorithm outperforms the prevailing ones.