Asbestos-containing materials pose a negative impact on human health and in environment sustainability in overall. Therefore, emergent actions should be taken for its removal in a best possible way. In this regard, Remote Sensing (RS) technology is an efficient tool for identification, mapping and monitoring of asbestos-containing (CA) roofs. Although the use of hyperspectral images from airborne sensors (e.g., MIVIS) has shown good results in achieving acceptable classification accuracy of CA roofs, some challenges have made such use unsuitable for large-scale mapping of these roofs, main-y due to the spatial resolution of hyperspectral images, which generates large errors of omission in detecting small CA roofs (less than 100 sq. m). In addition, in hyperspectral images, using an optimal number of bands is still a significant challenge due to the large number of bands and the resulting reduction in accuracy in terms of commissioning errors. Under this context , the research project employed and tested the potential of a specific U-Net architecture, as a deep learning (DL) convolutional neural network (CNN) model for automatic recognition of CA coverages. The metrics, evaluated on an initial U-Net neural network model, trained on ROIs (Region Of Interest) built only on simple RGB images, demonstrated a good ability of the network to generalize even complex pattern recognition (red roofs, industrial sheds, etc.), making any rotations, translations or small changes in the recognition process irrelevant. To identify discriminating "features" for asbestos roofs, the focus was on the extraction of statistical descriptors capable of detecting the regularity of Eternit corrugations, even for small roofs (less than 100 square meters). In order to characterize the specific textures related to AC roofs, in the research project, the use of the Gray-Level Co-occurrence Ma-trix (GLCM) was explored for calculating the frequency with which pairs of pixels occur with specific values and in a specific spatial relationship, thus calculating and testing numerous quantitative features and descriptors: mean value, moments, etc., related to AC coverages. The overall capability of the trained network led to excellent results in terms of accuracy, even finding values as high as 93% in industrial areas.
Rapporto tecnico - deliverable di progetto nazionale "MAAC Mappatura Amianto Ambiente Costruito" - Report specifiche tecniche e caratteristiche risultanti da pre-processamento per i dati di input relativi alla procedura di classificazione - Addestramento di una rete neurale
Antonietta Varasano
2022
Abstract
Asbestos-containing materials pose a negative impact on human health and in environment sustainability in overall. Therefore, emergent actions should be taken for its removal in a best possible way. In this regard, Remote Sensing (RS) technology is an efficient tool for identification, mapping and monitoring of asbestos-containing (CA) roofs. Although the use of hyperspectral images from airborne sensors (e.g., MIVIS) has shown good results in achieving acceptable classification accuracy of CA roofs, some challenges have made such use unsuitable for large-scale mapping of these roofs, main-y due to the spatial resolution of hyperspectral images, which generates large errors of omission in detecting small CA roofs (less than 100 sq. m). In addition, in hyperspectral images, using an optimal number of bands is still a significant challenge due to the large number of bands and the resulting reduction in accuracy in terms of commissioning errors. Under this context , the research project employed and tested the potential of a specific U-Net architecture, as a deep learning (DL) convolutional neural network (CNN) model for automatic recognition of CA coverages. The metrics, evaluated on an initial U-Net neural network model, trained on ROIs (Region Of Interest) built only on simple RGB images, demonstrated a good ability of the network to generalize even complex pattern recognition (red roofs, industrial sheds, etc.), making any rotations, translations or small changes in the recognition process irrelevant. To identify discriminating "features" for asbestos roofs, the focus was on the extraction of statistical descriptors capable of detecting the regularity of Eternit corrugations, even for small roofs (less than 100 square meters). In order to characterize the specific textures related to AC roofs, in the research project, the use of the Gray-Level Co-occurrence Ma-trix (GLCM) was explored for calculating the frequency with which pairs of pixels occur with specific values and in a specific spatial relationship, thus calculating and testing numerous quantitative features and descriptors: mean value, moments, etc., related to AC coverages. The overall capability of the trained network led to excellent results in terms of accuracy, even finding values as high as 93% in industrial areas.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
