Uporaba večmodalnih podatkov za zaznavanja terasiranih pokrajin z globokimi modeli

KORAČIN, MATIC (author), Čehovin Zajc, Luka (mentor), Ciglič, Rok (co-mentor)

Terasirane pokrajine so svetovno prepoznane kot pomemben del naravne in kulturne dediščine. V Sloveniji se je obdelovanje velikega dela teras opustilo, zato gotov zgodovinski obseg terasiranih pokrajin na ozemlju Slovencev ni natanko znan. Metode za detekcijo teras že obstajajo in temeljijo na semantični segmentaciji javno dostopnih višinskih podatkov, pridobljenih s tehnologijo LIDAR. V raziskavi smo preverjali vplive več vhodnih modalnosti kot je ortofoto, dejanska raba površja in meje med parcelami. Podatke smo obdelovali z variacijo arhitekture globokega učenja U-Net ter za združevanje podatkov primerjali tehniki zgodnje in srednje fuzije podatkov. Naučen napovedni model se je naučil locirati terasirane pokrajine bolje kot model iz referenčne raziskave. Naši rezultati kažejo, da uporaba večmodalnosti ne doprinese bistvenih dodatnih informacij pri zaznavanju terasiranih pokrajin.

• Language: Slovenian
• Keywords: terase, strojno učenje, segmentacija, večmodalnost, digitalni model višin, ortofoto
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-150186
• COBISS.SI-ID: 168521475
• Publication date in RUL: 14.09.2023

Secondary language

• Language: English
• Title: Using multimodal data for terraced land recognition with deep models

Abstract

Terraced landscapes are globally recognized as part of natural and cultural heritage. The maintenance of a large part of terraces in Slovenia has been given up, so the historical scope of terraced landscapes in Slovenia is not exactly known. Methods for landscaped areas detection have already been implemented and are based on semantic segmentation of elevation data extracted with LIDAR technology. In this research we examined the effect of using other modalities as input such as orthofoto, land usage data and borders between real estates. We processed the data with a variation of the U-Net architecure and compared early and middle data fusion techniques. The finished model has shown to be more effective at detecting terraced landscapes than the model in the reference research. Our results show that using extra modalities does not contribute significantly to the detection of terraced landscapes.

• Keywords: terraces, machine learning, segmentation, multimodality, digital elevation model, orthophoto