A transferable remote sensing approach to classify building structural types for seismic risk analyses: the case of Val d'Agri area (Italy)

Abstract

This study proposes a methodology based on machine learning (ML) algorithms for rapid and robust classification of building structural types (STs) in multispectral remote sensing imagery aiming to assess buildings’ seismic vulnerability. The seismic behavior of buildings is strongly affected by the ST, including material, age, height, and other main structural features. Previous works deployed in situ data integrated with remote sensing information to statistically infer STs through supervised ML methods. We propose a transferable methodology with specific focus on situations with imbalanced in situ data (i.e., the number of available labeled samples for model learning differs largely between different STs). We learn a transferable model by selecting features from an exhaustive set. The transferability relies on deploying geometric features characterizing individual buildings; thus, the model is less sensitive to domain adaption problems frequently induced by e.g., changes in acquisition parameters of remotely sensed imagery. Thereby, we show that few geometry features enable generalization capabilities similar to models learned with a large number of features describing spectral, geometrical or contextual building properties. We rely on an extensive geodatabase containing almost 18,000 building footprints. We follow a Random Forest (RF)-based feature selection strategy to objectively identify most valuable features for prediction. Furthermore, the problem of unbalanced classes is addressed by adopting two approaches: downsampling the majority class and modifying the classifier internally (weighted RF). The implemented model is transferred on the challenging urban morphology of the Val d’Agri area (Italy). Results confirm the statistical robustness of the model and the importance of the geometry features, allowing for reliable identification of STs.

Appendix

Features	Description	References: studies on characterization of built environments
Geometry-extent 2D
Area (IB_AREA; AB_AREA)	IB and AB polygon area	Steiniger et al. (2008), Hermosilla et al. (2014), Voltersen et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Length (IB_LENGTH; AB_LENGTH)	Length of the longer side of the minimum bounding rectangle comprising IB and AB polygons	Geiß et al. (2015) and Wurm et al. (2016)
Width (IB_WIDTH; AB_WIDTH)	Length of the shorter side of the minimum bounding rectangle comprising IB and AB polygons	Geiß et al. (2015) and Wurm et al. (2016)
Length/width (IB_L_W; AB_L_W)	Ratio between length and width of each IB and AB polygon	Steiniger et al. (2008), Voltersen et al. (2014), Geiß et al. (2015 and Wurm et al. (2016)
Perimeter (IB_PERIMET; AB_PERIMET)	IB and AB polygon perimeter	Hermosilla et al. (2014), Voltersen et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Inverted floor area ratio (IB_IFAR)	${\text{IB}}\, = \,\frac{polygon \,\,area}{polygon \,\,area \;\times\;number \,\,of \,floors}$	Berger et al. (2013)
BD Area (BD_AREA)	Area of each BD polygon	Steiniger et al. (2008)
Average built-up area (AV_IB_AREA)	Average area of the IB polygons within each BD polygon	Yu et al. (2010)
Sum built-up area (SUM_IB_AREA)	Sum of the IB polygon areas within each BD polygon	Steiniger et al. (2008)
Average L/W IB (AV_IB_L_W)	Average length/width based on IB polygons within each BD polygon	Voltersen et al. (2014)
Average L/W AB (AV_AB_L_W)	Average length/width based on AB polygons within each BD polygon	Voltersen et al. (2014)
Geometry-shape 2D
Number of building vertices (building corners) (IB_N_VERT, AB_N_VERT)	Count of polygon points of the exterior ring	Steiniger et al. (2008) and Wurm et al. (2016)
Shape index (IB_SI, AB_SI)	Describes the smoothness of the outer shape of the polygon object. Calculation is based on the proportion between the real perimeter P of the polygon and an approximated square with the same area as the polygon $SI = \frac{P}{4 \times \sqrt A }$	Belgiu et al. (2014), Hermosilla et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Fractal dimension (IB_FRAC_DI, AB_FRAC_DI)	Provides a numerical description of the complexity and segmentation of a polygon by computing the proportion of area and perimeter $FD = \frac{{{ \ln }\left( {\frac{U}{4}} \right)^{2} }}{lnA}$	Hermosilla et al. (2014) and Wurm et al. (2016)
Density (IB_DENS, AB_DENS)	The distribution in space of the pixels of an image object, calculated by the number of pixels forming the image object divided by its approximated radius, based on the covariance matrix. (Trimble 2014)	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Border index (IB_BI, AB_BI)	$BI = \frac{perimeter}{{2 \,\times\, \left( {length\, +\, width} \right)}}$	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Compactness (IB_COMP, AB_COMP)	$COMP = \frac{length \;\times\; width}{number \,\,of \,pixels}$	Hermosilla et al. (2014) and Belgiu et al. (2014)
Rectangular fit (IB_RECT_FIT, AB_RECT_FIT)	The calculation is based on a rectangle with the same area as the image object. The proportions of the rectangle are equal to the proportions of the length to width of the image object. The area of the image object outside the rectangle is compared with the area inside the rectangle. (Trimble 2014)	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Radius of largest enclosing ellipse (IB_RLEE, AB_RLEE)	The Radius of Largest Enclosed Ellipse feature describes how similar an image object is to an ellipse. The calculation uses an ellipse with the same area as the object and based on the covariance matrix. This ellipse is scaled down until it is totally enclosed by the image object. The ratio of the radius of this largest enclosed ellipse to the radius of the original ellipse is returned as feature value. (Trimble 2014)	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Radius of smallest enclosing ellipse (IB_RSEE, AB_RSEE)	The Radius of Smallest Enclosing Ellipse feature describes how much the shape of an image object is similar to an ellipse. The calculation is based on an ellipse with the same area as the image object and based on the covariance matrix. This ellipse is enlarged until it encloses the image object in total. The ratio of the radius of this smallest enclosing ellipse to the radius of the original ellipse is returned as feature value. (Trimble 2014)	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Elliptic fit (IB_EL_FIT, AB_EL_FIT)	The Elliptic Fit feature describes how well an image object fits into an ellipse of similar size and proportions. While 0 indicates no fit, 1 indicates a perfect fit. The calculation is based on an ellipse with the same area as the selected image object. The proportions of the ellipse are equal to the length to the width of the image object. The area of the image object outside the ellipse is compared with the area inside the ellipse that is not filled by the image object	Belgiu et al. (2014), Geiß et al. (2015)
Roundness (IB_ROUND, AB_ROUND)	The Roundness feature describes how similar an image object is to an ellipse. It is calculated by the difference of the enclosing ellipse and the enclosed ellipse. The radius of the largest enclosed ellipse is subtracted from the radius of the smallest enclosing ellipse. (Trimble 2014)	Belgiu et al. (2014), Geiß et al. (2015) and Wurm et al. (2016)
Normalized perimeter index (IB_NPI, AB_NPI)	The NPI is the proportion of the perimeter of a circle with the same area as the building polygon ($PA_{circle} )$ with the same perimeter of the building object $nPI = \frac{{PA_{circle} }}{P}$	Wurm et al. (2016)
Normalized proximity index (IB_NPRI, AB_NPRI)	The proximity index is based on the calculation of Euclidian distances between single pixels of an object and the object center $nPrI = \frac{{PA_{circle} }}{{PA_{object} }}$ where the proximity of the circle is $PA_{circle} = \frac{2}{3} \times r_{{A_{circle} }}$ and $r_{{A_{circle} }}$ is the radius of the circle and for $P_{object}$ $P_{object} = \mathop \sum \limits_{j = 1}^{n} d_{j} \times \frac{1}{n}$ where $d$ is the Euclidian distance of the pixel to the object center	Wurm et al. (2016)
Normalized spin index (IB_NSI, AB_NSI)	The nSI (or moment of inertia) is similar to the proximity index, but the moment of inertia weights the extremities of the polygon higher: $nSI = \frac{{J_{{A_{circle} }} }}{{J_{{A_{object} }} }}$ the moment of inertia for the circle is: $J_{{A_{circle} }} = \frac{1}{2}r_{{A_{circle} }}^{2}$ and for the object: $J_{object} = \frac{1}{n} \times \mathop \sum \limits_{j = 1}^{n} d_{j}^{2}$ where $d_{j}^{2}$ equals the square Euclidian distance of the pixels to the object center	Wurm et al. (2016)
Areal asymmetry (IB_ASYM, AB_ASYM)	The asymmetry feature describes the relative length of an image object, compared to a regular polygon. An ellipse is approximated around a given image object, which can be expressed by the ratio of the lengths of its minor and the major axes. The feature value increases with this asymmetry	Wurm et al. (2016)
Geomtry-extent 3D
Building height (IB_HEIGHT)	For each IB polygon the height has been calculated as: Average interstorey height × number of floors × area	Hermosilla et al. (2014)
Maximum height AB (AB_Max_HEIGHT)	Building height of AB polygons calculated considering the maximum height value of the IB polygons forming each AB polygon	Hermosilla et al. (2014)
Area weighed maximum height AB (AB_WEIGHT_A)	Area-weighed height based on maximum heights of IB polygons within AB polygons	Hermosilla et al. (2014)
Building average height (AB_MEAN_H)	Building height of AB polygons calculated considering the average height value of the IB polygons forming each AB polygon	Hermosilla et al. (2014)
Area weighed mean height (AB_MEAN_W_A)	Area-weighed height based on the average height of IB polygons forming each AB polygon	Hermosilla et al. (2014)
Maximum height BD (BD_MAX_HEIGHT)	Maximum height value within BD polygon	Hermosilla et al. (2014)
Area-weighted maximum height BD (BD_WEIGHT_A)	Area-weighed height based on maximum heights of IB polygons within BD polygons	Hermosilla et al. (2014)
Mean of building height within BD (AV_IB_HEIGHT)	Average height of IB within BD polygons	Voltersen et al. (2014), Hermosilla et al. (2014) and Geiß et al. 2015
Area-weighted on mean heights BD (BD_AV_MA)	Area-weighed height based on average heights of IB polygons within BD polygons	Yu et al. (2010)
Building volume (IB_VOLUME)	The building volume of the IB is calculated as: $IB_{VOLUME} = average\,floor\,area\,\times\,average\,interstorey\,height\,\times\,number\,of\,floors$	Yu et al. (2010)
Maximum building volume based on maximum height (AB_VOL_MAX_H)	Maximum building volume of AB polygons within BD polygons based on maximum height of IB polygons	Hermosilla et al. (2014)
Mean building volume (AB_AV_VOL)	Average building volume of AB polygons within BD polygons based on maximum height of IB polygons	Hermosilla et al. (2014)
Geometry-shape 3D
IB Shape index 3D (SI_3D)	Describes the smoothness of the object in three dimension; it is calculated by the proportion of the real perimeter with the approximated perimeter of a cube with the same volume than the real object $SI_{3D} = \frac{P}{{4 \times \sqrt[3]{V}}}$	Wurm et al. (2016)
Shape index 3D_maxHeight (AB_SI3D_MAX_H)	Shape index 3D of AB polygons based on area-weighted maximum heights of IB	Wurm et al. (2016)
Shape index 3D_meanHeight (AB_SI3D_AV_H)	Shape index 3D of AB polygons based on area-weighed mean heights of IB	Wurm et al. (2016)
Spatial context/configuration
Orientation (IB_ORIENT, AB_ORIENT)	Orientation of the major axis of the minimum bounding rectangle	Belgiu et al. (2014) and Geiß et al. (2015)
Distance to nearest building (IB_DIST, AB_DIST)	The shortest distance between buildings	** and comprehensive feature calculation for an automated derivation of urban structure types at block level. Remote Sens Environ 154:192–201" href="/article/10.1007/s10518-019-00648-7#ref-CR57" id="ref-link-section-d80340767e5878">2014)
Building aggregation measure (BD_BA)	$BA = \frac{{\frac{{A_{b} }}{{A_{AOI} }}}}{{Median\left( {D_{b} } \right)}} - \frac{{Median\left( {IFAR} \right)}}{{N_{B} }}$ where $A_{b}$, floor area covered by buildings; $A_{AOI}$, area of interest; $D_{b}$, distance to the nearest building; $IFAR$, Inverted Floor area Ratio; $N_{B}$, number of buildings	Berger et al. (2013)
Normalized building aggregation measure (BD_NBA)	${\text{NBA}} = \frac{{BA_{i} - BA_{min} }}{{BA_{max} - BA_{min} }}$ where $BA_{i}$, BA in the i-th BD polygon; $BA_{max}$, maximum BA value; $BA_{min}$, minimum BA value	Berger et al. (2013)
Spectral-1st order
Mean blue (IB_sMeanB, AB_sMeanB)	Mean intensity in the blue channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
Mean green (IB_sMeanG, AB_sMeanG)	Mean intensity in the green channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
Mean red (IB_sMeanR, AB_sMeanR)	Mean intensity in the red channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
Mean NIR (IB_sMeanNIR, AB_ sMeanNIR)	Mean intensity in the NIR channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
St. Dev. blue (IB_sSTDV_B, AB_ sSTDV_B)	Standard deviation of the intensity in the blue channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
St. Dev. green (IB_sSTDV_G, AB_ sSTDV_B)	Standard deviation of the intensity in the green channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
St. Dev. red (IB_sSTDV_R, AB_ sSTDV_R)	Standard deviation of the intensity in the red channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
St. Dev. NIR (IB_sSTDV_NIR, AB_sSTDV_NIR)	Standard deviation of the intensity in the NIR channel	Bruzzone and Carlin (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
Blue/green (IB_sB/G, AB_ sB/G)	Mean blue/mean green	Bruzzone and Carli (2006), Geiß et al. (2015) and Leinenkugel et al. (2011)
Blue/red (IB_sB/R, AB_ sB/R)	Mean blue/mean red	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Blue/NIR (IB_sB/NIR, AB_ sB/NIR)	Mean blue/mean NIR	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Green/red (IB_sG/R, AB_ sG/R)	Mean green/mean red	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Green/NIR (IB_sG/NIR, AB_ sG/NIR)	Mean green/Mean NIR	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Red/NIR (IB_sR/NIR, AB_ sR/NIR)	Mean red/mean green	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Normalized green (IB_sG/GRNIR, AB_ sG/GRNIR)	Mean(green)/[mean(green) + mean(red) + mean(NIR)]	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Normalized red (IB_sR/GRNIR, AB_ sR/GRNIR)	Mean(red)/[mean(green) + mean(red) + mean(NIR)]	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Normalized NIR (IB_sNIR/GRNIR, AB_ sNIR/GRNIR)	Mean(NIR)/[mean(green) + mean(red) + mean(NIR)]	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Brightness (IB_sBRIGHT, AB_ sBRIGHT)	Mean intensity of all channels	Bruzzone and Carlin (2006), Geiß et al. (2015) and Voltersen et al. (2014)
Normalized differenced vegetation index (IB_sNDVI, AB_ sNDVI)	$NDVI = \frac{NIR - red}{NIR + red}$	Geiß et al. (2015) and Leinenkugel et al. (2011)
Soil-adjusted vegetation index (IB_sSAVI, AB_ sSAVI)	$SAVI = \frac{NIR - red}{NIR + red + L} \times \left( {1 + L} \right)$ where L = 0.5	Bruzzone and Carlin (2006) and Geiß et al. (2015)
Spectral-2nd order
$GLCM_{inv.}$ (Angular 2nd moment) (IB_GLCM_ANG, AB_GLCM_ANG)	$\mathop \sum \limits_{i,j = 0}^{N - 1} \left( {Pi,j} \right)^{2}$ where i, row number; j, column number; $P_{i,j}$, normalized value in the cell i, j; N is the number of rows or columns	Geiß et al. (2015) and Zhang et al. (2006)
$GLCM_{inv.}$ (entropy) (IB_GLCM_ENT, AB_GLCM_ENT)	$\mathop \sum \limits_{i, j = 0}^{N - 1} P_{i,j} ( - \ln P_{i,j} )$ where i, row number; j, column number; $P_{i,j}$, normalized value in the cell i, j; N, number of rows or columns	Geiß et al. (2015) and Zhang et al. (2006)
$GLCM_{inv.}$ (homogeneity) (IB_GLCM_HOM, AB_GLCM_HOM)	$\mathop \sum \limits_{i, j = 0}^{N - 1} \frac{{P_{i,j} }}{{1 + \left( {i - j} \right)^{2} }}$ where i, row number; j, column number; $P_{i,j}$, normalized value in the cell i, j; N, number of rows or columns	Geiß et al. (2015) and Zhang et al. (2006)

1. List of the state-of-the-art features deployed to describe the building stock in the present study