Benefits of PRISMA hyperspectral data for tree species classification in an area of high forest biodiversity

Abstract

The Serre Regional Park in the south of Italy (lat. 38.55, long. 16.35) extends over an area of 17900 hectares, largely covered by forests with a high level of biodiversity. The forest cover is mainly composed of about 15 tree species. This type of area is a natural laboratory for experimenting with forest classification techniques. Indeed, this work aimed to test how to use remote sensing (RS) hyperspectral data combined with innovative AI-based classification techniques to classify forest tree species. The potential of RS for monitoring agricultural and forestry conditions is enhanced by the wealth of information provided by hyperspectral imagery (HSI) and new classification techniques. HSI derived from recent satellite missions (e.g. PRISMA or EnMAP) provides information in multiple bands, from visible/near infrared (400-1010nm) to shortwave infrared (920-2505nm). In addition, the last decade has seen renewed interest in Deep Learning (DL) methods. In particular, convolutional neural networks (CNN) have been widely used in the analysis of images.The study area includes the Park and corresponds to an acquisition of 900 km2 of the PRISMA satellite (taken in July 2023). The PRISMA L2D level cube (Cube 1) used in this study was first processed, for format conversion and georeferencing improvement of the image, with a Python script developed by the authors (www.github.com/LarpUnina/PrismaTool). Next, two different techniques were used to reduce the dimensionality. A second cube (Cube 2) was obtained using a band selection operation and a third cube (Cube 3) was obtained using a PCA technique. For the next classification step, both cubes were used as input. Specifically, a Convolutional Neural Network was selected to classify the data using the open source AVHYAS plugin in the Qgis environment. Ground truths were derived from four SAC site plans provided by the Park Authority, covering approximately 9000 hectares, and were split (70 - 30 %) both to train the network and to test the results. The classes chosen for the classification task includes the most common tree species in the Park area: chestnut (Castanea sativa), larch pine (Pinus nigra), alder (Alnus glutinosa), beech (Fagus sylvatica), silver fir (Abies alba), oak (Quercus ilex) and poplar (Populus alba).  Cube 2 gave better results than Cube 3 as input data with an OA higher than 90%. The best results with F1 around 90% among tree species were obtained for Fagus sylvatica, Abies alba and Castanea sativa. Populus alba was the species with less accurate results. HSI contributes to a better definition of the trends of the spectral signature of trees and makes it possible to distinguish even between similar species.