Day 4

Detailed paper information

Back to list

Paper title Forest species mapping in India using semi-supervised learning of hyperspectral images
  1. Debmita Bandyopadhyay Cambridge University Speaker
  2. Madeleine Kotzagiannidis
  3. Carola Schönlieb Cambridge University
  4. David Anthony Coomes Cambridge University Speaker
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text Forest managers are increasingly interested in monitoring forest species in the context of conservation and land use planning. Field monitoring of the dense tropical forests is an arduous task, so remote sensing of tree species in these regions poses a great advantage. Hyperspectral imaging (HSI) offers a rich source of information, comprising reflectance measurements in hundreds of contiguous bands, making it valuable for image classification. Many pixel-based algorithms have been used in image classification, such as support vector machines (Melgani and Bruzzone, 2004), neural networks (Ratle et al., 2010), open learning (Li et al., 2011), to name a few. However, these approaches are strongly dependent on the dimensionality of the data and require many more labelled samples than are typically available from field surveys. The latter is usually challenging to obtain as they are based on data manually collected data on the ground.

To circumvent the problem of having few labels, in this study, we show how a semi-supervised spectral graph learning (SGL) algorithm (developed by Kotzagiannidis and Schönlieb in 2021 on standard HSI dataset), in conjunction with superpixel clustering, can be used for forest species classification. This new approach is based on three main steps: 1) the SLIC segmentation algorithm that creates superpixels considering both the size and resolution of the HSI image 2) using the label propagation on nearest neighbouring superpixels an initial smooth graph is learnt based on the features extracted in the image, and 3) the learnt graph is updated utilizing penalizing functions for classes not belonging to the class, followed by label propagation and the final class assignment. We used this new approach to classify the tropical forest species from airborne hyperspectral imagery collected by NASA’s AVIRIS sensor in the Shivamogga forested region of southern India. In the surveyed area we labelled tree crowns of 31 tree species, of which three species - Terminalia tomentosa, Terminalia bellirica and Anogeissus latifolia - were labelled over ten times. It is to note that only 5% of the data under consideration had labels, still, the SGL method improved in performance (2%) compared to the linear graph learning (Sellars et al.,2020) but substantially better than Support Vector Machine algorithm (11%), Local Global Consistency (9%), based on the Kappa coefficients.

The main reason for the better performance of SGL over other approaches is the incorporation of multiple features into the updatable graph. This approach refines the graph to the extent that it can capture the complex dependencies in the HSI data and ultimately provide an improved classification performance. With the method now tested in complex mixed tropical forests using AVIRIS hyperspectral images, this state-of-art algorithm looks promising for application in forests in other regions of the world.

Kotzagiannidis MS, Schonlieb CB. Semi-Supervised Superpixel-Based Multi-Feature Graph Learning for Hyperspectral Image Data. IEEE Trans Geosci Remote Sens 2021.
Li J, Bioucas-Dias JM, Plaza A. Hyperspectral image segmentation using a new Bayesian approach with active learning. IEEE Trans Geosci Remote Sens 2011;49:3947–60.
Melgani F, Bruzzone L. Classification of hyperspectral remote sensing images with support vector machines. IEEE Trans Geosci Remote Sens 2004;42:1778–90.
Ratle F, Camps-Valls G, Weston J. Semisupervised neural networks for efficient hyperspectral image classification. IEEE Trans Geosci Remote Sens 2010;48:2271–82.
Sellars P, Aviles-Rivero AI, Schonlieb CB. Superpixel Contracted Graph-Based Learning for Hyperspectral Image Classification. IEEE Trans Geosci Remote Sens 2020;58:4180–93.