doktorska disertacija
Andrej Kobler (Author), Dušan Kogoj (Mentor), Krištof Oštir (Mentor), Matjaž Mikoš (Thesis defence commission member), Mojca Kosmatin Fras (Thesis defence commission member), Tatjana Veljanovski (Thesis defence commission member), Marko Debeljak (Thesis defence commission member), Sašo Džeroski (Co-mentor)

Abstract

Naloga je sestavljena iz dveh sklopov. V prvem sklopu smo razvili novo metodo izdelave lidarskega DMR, v drugem pa smo lidarske vertikalne profile vegetacij uporabili za modelno napovedovanje deležev drevesnih vrst v gozdu in svetlobnih lastnosti gozda. Obstoječi algoritmi za izračun DMR iz lidarskih podatkov imajo v strmem gozdnatem reliefu težave z razlikovanjem talnih odbojev od vegetacije, saj ima v strmem reliefu lokalna okolica v oblaku podobne značilnosti kot vegetacija. V prvem sklopu naloge smo razvili novo metodo izdelave DMR, imenovano REIN, posebej namenjeno uporabi v strmem gozdnatem reliefu: Metoda za zmanjšanje napak izkorišča preobilje lidarskih odbojev (redundanco) in ne spada v nobeno od skupin znanih algoritmov, saj uporablja naključno vzorčenje oblaka lidarskih odbojev. REIN ima večjo zmožnost prilagajanja raznolikim razmeram v strmem gozdnatem terenu. Ker ima vsak del analiziranega območja enako možnost biti izbran v vzorec, je izračunan DMR homogen tudi pri nehomogenih vhodnih podatkih. REIN rešuje tudi problem negativnih osamelcev, ki so rezultat mnogokratnih odbojev. V drugem sklopu naloge smo lidarske vertikalne profile vegetacije, izračunane iz podatkov diskretnega lidarja majhnega odtisa, uporabili za modelno napovedovanje deležev drevesnih vrst v gozdu in svetlobnih lastnosti gozda. Za gradnjo modelov smo uporabili ansambelske metode strojnega učenja ter različne kombinacije pojasnjevalnih spremenljivk, izpeljanih iz diskretnih lidarskih podatkov in iz infrardečih aeroposnetkov. Za osem najbolje napovedanih ciljnih spremenljivk modelne korelacije znašajo od 0,76 do 0,83. Razmeroma nizke vrednosti pripisujemo raznolikosti gozda v testnem območju, napakam v učnem vzorcu in nenatančnemu pozicioniranju vzorčnih ploskev (nenatančen GPS v gozdu), zaradi česar je lahko prišlo do zamika terenskih glede na daljinsko zaznane podatke. Pri ciljnih spremenljivkah, ki se nanašajo na drevesno sestavo, k točnosti napovedi največ prispevajo infrardeče pojasnjevalne spremenljivke, lidarski podatki pa so boljši pri pojasnjevanju svetlobnih značilnosti gozda, ki so tesno povezane s prostorsko razporeditvijo biomase. Strojno naučeni ansambelski modeli so točnejši in robustnejši od lineranih regresijskih modelov, večciljni modeli pa so primernejši od modelov z eno ciljno spremenljivko, saj je skupen čas za pripravo enega modela krajši od časa za pripravo mnogih enociljnih modelov, pa tudi lažje jih je implementirati. Gozdni prostor smo razčlenili na gozdne sestoje s slikovno segmentacijo na podlagi modelnih rastrskih kart.

Keywords

geodezija;disertacije;lidar;gozd;DMR;strojno učenje;model;

Data

Language: Slovenian
Year of publishing:
Typology: 2.08 - Doctoral Dissertation
Organization: UL FGG - Faculty of Civil and Geodetic Engineering
Publisher: [A. Kobler]
UDC: 528.7/.8:630*5:(043.3)
COBISS: 5489249 Link will open in a new window
Views: 2125
Downloads: 634
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: New methods of prosessing [!] aerial laser scanner data for forest ecosystem monitoring
Secondary abstract: The thesis is composed of two parts. In the first part we developed a new method for lidar DTM generation. In the second part we used vertical lidar profiles for model-based prediction of the percentages of individual tree species in the forest and to predict different light properties of the forest In steep forested relief, the existing algorithms for computing DTM from the lidar data have problems to distinguish between the ground returns and the vegetation returns, because on the steep slopes the local cloud neighborhood has properties similar to the vegetation. In the first part of the thesis we introduced a new method of DTM computation from the lidar data, called REIN, which is especially adapted to the steep forested topography. The method makes use of the lidar point redundancy to mitigate errors. It does not belong into any of the known algorithm groups, because it randomly samples the point cloud. REIN has a greater ability to adapt to variations in the terrain and forest cover. Because of ensuring that each part of the area of interest gets equal probability of being sampled, REIN results in a homogeneous DTM even under non-homogeneous data input conditions. REIN also takes care of the problem of negative outliers due to multi-path reflections. In the second part of the thesis we used vertical vegetation profiles, computed from the small-footprint discrete lidar data, to predict the percentages of individual tree species in the forest and to predict different light properties of the forest. The ensemble methods of machine learning were used together with different combinations of the explanatory variables, derived both from the discrete lidar data and from the aerial infra-red imagery. The correlations for the eight best modeled target variables are between 0,76 and 0,83. Relatively modest correlations are attributed to the heterogeneity of forests in the test area, to the errors in the training set, and to the imprecise positioning of the field plots (due to GPS errors under the forest canopy), resulting in a possible spatial shift between the field data and lidar data. Infrared explanatory variables contribute the most to the predictions of target variables referring to the tree composition. Lidar data are better suited to explain the forest light properties, which in turn are linked to the spatial distribution of the above-ground forest biomass. The machine-learned ensemble models are more accurate and more robust than the linear regression models. The multi-target models are more suitable than the single-target ensemble models, because the total time to set up a multi-target model is shorter than the time needed to set up multiple single-target models. The multi-target models are also easier to implement. The forest has been delimited into the forest stands by image segmentation based on the model-based raster maps.
Secondary keywords: doctoral thesis;deformation analysis;lidar;forest;DTM;machine learning;model;
File type: application/pdf
Type (COBISS): Dissertation
Thesis comment: Univ. v Ljubljani, Fak. za gradbeništvo in geodezijo
Pages: XVIII, 131 str.
ID: 8711006