Multispektralne podobe, bodisi letalski ali satelitski posnetki, zagotavljajo podrobne informacije o stanju površja. Klasifikacija na podlagi multispektralnih podob je ena izmed temeljnih nalog daljinskega zaznavanja. Podatke o stavbah vodimo v katastru stavb, ki ga moramo redno vzdrževati. Kot alternativne metode vzdrževanja se lahko uporabljajo metode računalniškega vida in strojnega učenja. Pri samodejni klasifikaciji stavb v zadnjih letih prednjačijo metode globokega učenja z uporabo konvolucijskih nevronskih mrež. V raziskavi smo uporabili konvolucijsko ogrodje Mask Region Convolutional Neural Network (Mask R-CNN) in razvili lastno podatkovno zbirko v formatu Microsoft Common Objects in COntext (MS COCO), ki smo jo uporabili za učenje modela zaznavanja stavb na podlagi bližnje infrardečih posnetkov cikličnega aerosnemanja (CAS) območja Slovenije iz leta 2019. Metodo za prepoznavo stavb smo preizkusili na izbranih območjih v Sloveniji in rezultate klasifikacije stavb analizirali. Rezultati samodejne klasifikacije stavb kažejo, da so metode globokega učenja primerne za iskanje in vzdrževanje podatkov o stavbah ter lahko nadomestijo ali se uporabijo kot pomoč pri že obstoječih metodah samodejne klasifikacije stavb.
| Language: | English |
|---|
| Title: | Automatic classification of buildings with deep learning : master thesis no.: 109/II. GIG |
|---|
| Abstract: |
|---|
Multispectral satellite or aerial images provide detailed information about the Earth%s surface. Multispectal image based classification is one of the fundamental tasks in the field of remote sensing. Building data is organized in the buidling cadastre, which needs to be regularly updated. As alternative methods for building cadastre maintenance computer vision and machine learning can be used. In recent years deep learning with the emphasis on convolutional neural networks are in the forefront for automatic classification of buildings. We applied the region based convolutional framework called Mask Region Based Convolutional Neural Network (Mask R-CNN) for automatic building classification and developed a dataset in the Microsoft Common Objects in Context (MS COCO) format. The building dataset was used for the training of the models on near infrared aerial images from the last aerial imaging of Slovenia in year 2019. The proposed method was tested and evaluated on selected areas in Slovenia. The results show that automatic classification of buildings with deep learning is suitable for building detection and can be used either as a replacement of current techniques or to aid the existing ones.
|
| Keywords: | geodesy, master thesis, deep learning, convolutional neural networks, classification of buildings, CAS, Mask R-CNN, object detection, object segmentation, automatic classification |
|---|
