Get More Info!

Announcement
Announcement
Classification and mapping of forest tree species using fast learning methods with satellite remote sensing data

Student name: Mr Akash Narayanan
Guide: Dr Ayushi Vijhani
Year of completion: 2024
Host Organisation: Panacea Geospatial Solutions Pvt Ltd.
Supervisor (Host Organisation): Dr Veeranarayana Balabathina
Abstract:

The identification and classification of individual forest tree species is a very important method and can provide an efficient and useful tool for the classification of forest and for forest management and helping in planning and monitoring processes. The EO-1 Hyperion satellite data which is a Hyperspectral data provide the useful and sufficient spectral information that is required for the classification of individual forest tree species. This research does a comparative analysis between 13 supervised classifiers and their accuracies and results in a complex forest which also has tree species distributed in a heterogenous pattern in the Hauz Khas Forest in Delhi, India. The 13 classifiers are Binary Encoding, Minimum Distance, Neural Net, Parallelepiped, Maximum Likelihood, Mahalanobis Distance, Spectral Angle Mapper, Spectral Information Divergence, Support Vector Machine, Random Forest Classifications, K-Nearest Neighbour, Decision Tree Classification and Artificial Neural Networks. The field data collection is also done to collect the training and testing samples, which is also analysed in this study. A pixel-based training samples is explored and the field data points for training and testing samples are trained accordingly. The research also experiments with four different types of datasets as well. The four datasets are the following i) the data with all the bad bands removed, the All-Bands dataset with 167 bands, ii) the Selected Bands data with 27 bands on the basis of objects and signatures-based band ratios, iii) the data dimensionality reduced, a Principal Component Analysis, with 4 bands based on Eigenvalues and iv) the Minimum Noise Fraction transform with 4 bands, based on the bands with the lowest noise and less disturbances. Each of the classifiers were assessed individually in all the separate datasets and to check if there are any advantages related to an increase in the training sample size. All the classification of different classifiers were giving accuracies, mostly poor accuracies of most classifiers like the Parallelepiped, but then a higher result in the accuracies was seen in all the datasets of Random Forest Classification.

Keywords: Hyperspectral Satellite Data, Binary Encoding, Minimum Distance, Neural Net, Parallelepiped, Maximum Likelihood, Mahalanobis Distance, Spectral Angle Mapper, Spectral Information Divergence, Support Vector Machine, Random Forest Classifications, K-Nearest Neighbour, Decision Tree Classification and Artificial Neural Networks, Principal Component Analysis, Minimum Noise Fraction.