Aplicación de técnicas de teledetección en ecosistemas macaronésicos

  1. Francesca Marchetti
Supervised by:
  1. Manuel Arbelo Pérez Director
  2. Bjorn Waske Director

Defence university: Universidad de La Laguna

Year of defence: 2021

Committee:
  1. Pedro Alberto Hernández-Leal Chair
  2. José Rafael García Lázaro Secretary
  3. Artur José Freire Gil Committee member

Type: Thesis

Abstract

The climatic and orographic conditions of Tenerife Island (Canaray Islands, Spain) have allowed the development of a rich biodiversity and different ecosystems. However, such a fragmented territory, the spectral similarity and complex spatial distribution of the species of vegetation, complicates the task of maintaining updated cartography using the traditional, costly and time-consuming fieldwork. Because of these limitations, this thesis proposed the use of remote sensing techniques from satellites and unmanned aerial vehicles in conjunction with selective field spectroradiometry, to map two important ecosystems situated in Tenerife: an agro-ecosystem dominated by chestnut trees (Castanea sativa Mill.) in the north slope of the island and a natural ecosystem (Malpaís de Güímar) in the southeast, with vegetation of beekeeping interest where the Cardonal-Tabaibal prevails. We performed thee different studies. In the first, the potential of very high- resolution (VHR) WorldView imagery and extended morphological profiles for mapping chestnut trees in the agroecosystem was analysed. Secondly, we spectrally characterized the vegetation of beekeeping interest in the Malpaís de Güímar and determined the wavelengths that best discriminated the vegetation species using a spectral separability analysis. Finally, we analysed the potential of 10 cm spatial resolution hyperspectral images to obtain cartography of the selected plant species in the Malpaís de Güímar. All remote sensing datasets (images) were classified with the Random Forest (RF) machine-learning algorithm to obtain the thematic maps of the ecosystems. RF is an adequate algorithm for mapping classes with complex characteristics, such as many of the plant species considered in this study. It demonstrated its ability to classify, with overall accuracies greater than 85%, both high-dimensional dataset (WorldView and hyperspectral images). In relation to the agro-ecosystem, two VHR WorldView images (March and May) to cover different phenological phases of chestnut trees were used. Moreover, was included spatial information in the classification process by extended morphological profiles (EMPs). The detailed accuracy assessment clearly reveals the benefits of multitemporal images in terms of mapping accuracy. The overall accuracies of the mono-temporal classifications are increased between 2% and 15%, when compared to the results achieved on the multitemporal data set. The inclusion of spatial information by EMPs further increases the classification accuracy by 5% and reduces the quantity and allocation disagreements in the final map. During years 2017 and 2018, an intense fieldwork was carried out to measure, with the ASD FieldSpec 3 spectroradiometer, the characteristic spectral signatures of the plant species with the greatest pollination potential in the Malpaís de Güímar. A spectral separability analysis, with the Jeffries-Matusita distance, made it possible to find the fourteen best spectral bands to discriminate the species. This analysis was customized for a Resonon Pika L camera that captures hyperspectral data from 400 nm to 1000 nm from a UAV. The results show that only eleven plant species show spectral separability with values higher than 1.9. The red-edge interval (705.5 nm - 757.5 nm) is particularly noteworthy. For the mapping of the Malpais de Güímar, a hyperspectral image covering approximately 6 ha was selected. Eight thematic classes were identified and defined: Aulaga, Balo, Cardón, Salado, Tabaiba amarga, Suelo, Sombra and Colmenas. From the original image of 120 spectral bands (ORIGINAL), two other reduced dimension datasets were constructed: PCA, which included the first five principal components; and SPECTRAL with the fourteen spectral bands of the previous spectral separability analysis. The ORIGINAL and PCA images showed the best results, with overall accuracies of 91.5% and 91.3%, respectively. However, no significant differences were found between them. The best-classified plant species was Cardón with a commission error of 1.1% and an omission error of 4%. In general, the thesis shows a methodology to generate accurate maps and monitor changes in the two relevant macaronesic ecosystems under study. The maps obtained in this work, increase and improve the information available so far, and could help reduce the vulnerability of these ecosystems to global climate change, strengthen their adaptation processes to prevent the loss of biological diversity and promote the development of sustainable activities to achieve their conservation.