New Digital Field of Drawing and Survey for the Automatic Identification of Debris Accumulation in Flooded Areas

Abstract

Floods are among the most devastating natural hazards in the world, causing the largest percentage of deaths and property damage. The impact of floods can be mitigated with an adequate knowledge of the territory, which makes it possible to better organize prevention plans with an appropriate analysis of the risk areas, which allows the management of relief efforts quickly and adequately. This work presents a methodology for mapping and monitoring the areas affected by floods and landslides by remote sensing: the correct representation and full interpretation of the territory matrix are essential for quality and sustainability design. In this paper, we used tools and technology that allow us to analyze and visualize the landscape evolution. The calibration of the method was performed on the events that took place in Calabria, in Southern Italy, on 12 August 2015. The proposed methodology concerned the planning of acquisition flights, the selection and setting of the sensors used, and the processing and post-processing of the data collected through the application of algorithms used for data manipulation and interpretation. The process of recognizing the areas with deposits of debris through the spectral signature was carried out using similarity criteria relating to hyperspectral data. The results obtained recommend the adoption of this methodology to deal with emergencies due to flood events.

1. Introduction

Among the most frequent natural events in recent decades, floods have undoubtedly recorded the most damaging consequences for people, the territory, and economic and social assets [1]. The risk of flooding is expected to increase, as is the number of people potentially exposed due to climate change and the conurbation. It is not clear whether the major consequences of flooding in recent decades are caused by more frequent and more intense floods or by the increased vulnerability of coastal areas and floodplains which are the preferred locations for settlement and economic development. Climate change is responsible for an increase in the probability of occurrence of floods and an increase in the magnitude of the danger, while it is certain that demographic and economic development cause a continuous increase in the vulnerability of many flood areas [2].

The geological history and the geographic, meteorological, and morphological conditions of the Italian peninsula determine a context of varied and widespread natural risk throughout the territory. Italy is subject to various natural disasters, from seismic to hydrogeological risks, from tsunamis to forest fires and interfaces, to end with the risk of active volcanoes, some of which are associated with a potentially devastating risk [3]. The phenomena related to meteorological and climatic forcings within this framework certainly represent a topical issue. Hydrogeological instability represents an extraordinarily significant problem in terms of impact on the population, on linear and service infrastructures, the economic and productive fabric, and cultural heritage [4,5]. Floods are natural phenomena that it is impossible to prevent. However, some human activities and climate change contribute to increasing the probability and aggravating negative impacts. During floods, rivers occupy areas of their own pertinence, that is, areas intrinsically connected to the natural evolution of the watercourse but which, over the years, have been occupied by the built environment. Thus, the phenomenon of flooding is often improperly defined as a natural disaster, forgetting that it is a natural event that is defined as calamitous as it affects man and his activities. In Italy, the exposure to flood risk is particularly high and constitutes a problem of social importance both for the number of victims and for the damage caused to buildings, industries, and infrastructures.

For correct management of the territory, it is, therefore, necessary to carry out an assessment of the hydrogeological risk. The risk indicates the potential losses associated with an extreme event that occurs in each area and with a given return period and can be defined in terms of negative consequences and probability of occurrence [6]. Risk consists of three factors: Hazard (H), Vulnerability (V) and Exposure (E). With H, we indicate the probability of occurrence of a flood event in a predetermined time interval and in a certain area. With V, we indicate the degree of loss for a given element at risk or a combination of elements subject to an event with a given intensity. Finally, with E, we indicate the extent of the elements at risk, such as people, goods, or property, which can be damaged when an event occurs, measured differently depending on their nature. The product of the three components yields a risk assessment [7].

The Directive 2007/60/EC of the European Parliament of 23 October 2007 [8], relating to the assessment and management of flood risk, requires the assessment of the risk through hydraulic studies. In this context, there is a need to use predictive tools and analysis methodologies suitable for determining the real risk conditions associated with the occurrence of a specific flood event. A flood study in an urban area must necessarily involve a series of aspects: to be of real support for the implementation of adequate forecasting and prevention measures. They range from the acquisition of topographical data to the description of the phenomenological processes typical of the motion of a current in flood and its interaction with buildings and infrastructures, up to the algorithms to be used for the solution of the model equations to the return of the results [9]. These analyses must lead to a final representation of the results through graphic products, the so-called hazard and risk maps. In this sense, the choice of a correct “point of view” is fundamental because the act of placing one’s eyes on something contains the choice by the narrator (here, represented in the meaning of the one who draws) to place his knowledge in a precise point from which to observe the reality. The choice of the “point of view” from which to observe and represent the territory is based on the synthesis of the multidimensionality of the features to be represented, on the different representation methodologies able to describe them, and on the reason for the representation itself, with the information that this must suggest, especially important in the event of environmental emergencies.

In addition, risk communication is of crucial importance, aimed at involving the various stakeholders and implementing proper emergency management [10]. It is, therefore, a question of returning the results in images that can immediately represent the conditions in which a place would be found during a flood in the form of virtual scenarios. Since the images must visually represent the calculated water surfaces, the procedure must necessarily be based on calculation models that correctly describe the hydraulic phenomena occurring in flood currents [11].

However, these calculation models must necessarily be validated by observations of real conditions. In this context, Earth observation techniques have become essential tools for managing this type of risk, both in the damage mapping phase following calamitous events and in the previous and subsequent phases of mitigation and restoration. In fact, remote sensing allows the identification of flooded areas and provides valuable information on the intensity, evolution of flooding and any damage to embankments or defense works [12]. The images detected with hyperspectral sensors, therefore, constitute the most efficient and rapid means for collecting data useful for ex-ante, in itinere, or ex-post analysis. Remote sensing is defined as the set of observation methods and techniques which, by extending the perceptive capacity of the human eye, provide qualitative and quantitative information on objects placed at a distance by measuring the electromagnetic radiation emitted or reflected by surfaces [13]. These techniques allow the environmental monitoring of areas whose extensions make a synchronous and spatially distributed characterization from land impracticable. Monitoring can be performed at different distances, ranging from meters in proximal sensing to kilometers in remote sensing [14,15].

Dewan et al. [16] performed a flood risk assessment in Bangladesh using remote sensing and GIS techniques. The authors used synthetic aperture radar (SAR) data by exploiting the hydrological parameters, frequency and depth of the flood calculated from the SAR images. In addition, elevation, land cover classification, geomorphic division and drainage network data generated by optical remote sensing and analog maps were used. Gaurav et al. [17] analyzed the 2010 Indus flood in Pakistan using remote sensing. The event was one of the largest river disasters in recent history, affecting more than 14 million people in Pakistan. The authors carried out an analysis of the accumulation of flow through SRTM (Shuttle Radar Topography Mission) data, highlighting the long stretches of the Indus River at risk of flooding. The method makes it possible to identify areas at risk and plan mitigation measures. Patel et al. [18] analyzed the flood risk using remote sensing data and GIS and made some correspondence with the urban planning scheme for some areas of India. The authors used high-resolution Google Earth images, images returned by the IRS-1D satellite, and 1: 50,000 topographic maps combined with the digital elevation model (DEM) to identify the area susceptible to flooding. The results indicate that more than 90–95% of the analyzed area would be submerged if a flood of the same frequency occurred on the same floodplain. Qi et al. [19] have developed a decision support system based on remote sensing data and GIS for the integrated management of floods in conditions of uncertainty through two-dimensional numerical simulations. The system leverages Remote Sensing (RS) imagery and GIS capabilities for flood damage calculations and loss of life estimates. The method also uses the Monte Carlo simulation to consider the uncertainties for estimating the population dynamics. Zhang et al. [20] used remote sensing to monitor and assess flood disasters in China. The authors tested a system for monitoring and assessing flood disasters that uses remote sensing, geographic information systems, and the global positioning system. The system was crucial for flood mitigation during the trial period by becoming an essential part of the flood management system at the Chinese flood control headquarters. Tien Bui et al. [21] exploited remote sensing and GIS for alluvial spatial modeling in northern Iran. The authors evaluated the performance of a model based on the Evidential Belief Function (EBF), also in combination with logistic regression (LR) methods, for the preparation of flood susceptibility maps for a catchment area. The EBF model returned the highest accuracy in forecasting susceptibility to flooding within the basin. The results obtained represent valid support for the planning and management of areas vulnerable to floods to prevent damage caused by floods. Hashemi-Beni et al. [22] adopted an approach based on Deep Learning and remote sensing data with a UAV for mapping the extent of a flooded area. Le Convolutional Neural Network (CNN) [23] is used to extract flooded areas from optical images. The increase in data is used to improve the results of the classification. The integrated method efficiently detects floods in both visible and overgrown areas. Samanta et al. [24] used remote sensing for the analysis of the susceptibility to flooding of the territory of Papua New Guinea (PNG). The authors developed a frequency ratio (FR) based model to manage several independent variables using weighted bivariate probability values to generate a plausible flood susceptibility map. Chen et al. [25] used remote sensing data to identify the impact of floods on agricultural production in northeastern China. The authors plotted the Enhanced Vegetation Index (EVI) during the growing seasons of the crops to compare them and attempt to detect and remove interference from natural intra-annual crop variations. The results obtained showed that this methodology is useful for detecting flood disturbances to crops and facilitating informed decision-making in agricultural flood management.

In this study, maps extracted from national datasets for hydrogeological risk management were compared with maps obtained from remote sensing activities carried out in situ. Following a flood event that affected a large area in the northeast of the Calabria Region in southern Italy, a large study was conducted on the use of airborne remote sensing for the identification and subsequent estimation of damage to services and infrastructure. Given the variety of phenomena found on the ground in the times following the flood event, the research team opted for the use of the airborne multi-sensor platform consisting of an Itres CASI1500 hyper-spectral chamber complete with INS inertial platform and GPS antenna. The integration of the sensors with specific flight plans and the comparative reading of the respective data on specific software platforms, according to a scientific protocol developed by the researchers, opens multiple thematic elaborations aimed at discretizing the complexity of anthropic and natural phenomena not easily detectable with the naked eye.

The rest of the paper is structured as follows: Section 2 describes in detail the multi-sensor airborne platform with which the surveys were carried out and the methods applied for the elaboration and subsequent analysis of the thematic maps and for the study of the forecasting models. Section 3 reports the results obtained by comparing the maps obtained from the remote sensing data with those obtained from the national datasets for the management of hydrogeological risk. In Section 4, the conclusions are drawn by proposing possible practical uses of the proposed technology and possible future research ideas.

2. Materials and Methods

2.1. Environmental Monitoring by Remote Sensing

Remote sensing is a technology that finds application in numerous contexts. It plays a primary role in environmental monitoring. It is a science that allows us to identify, measure, and analyze the qualitative and quantitative characteristics of a given object, area, or phenomenon without coming into direct contact with it [26]. By operating from above, at different distances and at different times, this discipline has introduced a new philosophy of control and investigation in the study of the territory and the related problems, allowing us to observe phenomena that are not directly accessible and, therefore, to overcome the difficulties associated with ground measurement campaigns, such as great organizational efforts, time, and resources not always available [27]. In recent years, the resolutions, both geometric and spectral, of the sensors used have significantly improved, allowing applications to be extended to the management of natural disasters, where the possibility of investigating the Earth’s surface without meeting it takes on a particular relevance, in consideration of the limited accessibility of the areas affected by a catastrophe. The impact of disasters, in fact, can only be mitigated with adequate knowledge of the territory. This knowledge makes it possible to better organize prevention plans through the assessment and determination of risk areas and plan to implement and manage relief quickly and adequately.

Without disturbing the observed object, remote sensing has the advantage of obtaining information about:

• Synoptics, that is, you get a high spatial coverage and instant observation of large areas.
• Dynamics, as it is possible to acquire data at different times in the same area of interest.
• Homogeneity, as there is consistent data.

Data acquisition takes place using appropriate sensors housed on aircraft that measure the flow of energy associated with the electromagnetic radiation that the surfaces emit or reflect. The electromagnetic waves emitted by a body have wavelengths that depend on the temperature of the body itself and on the physical, chemical, and geometric characteristics of its surface [28]. If we consider an electromagnetic wave incident on a surface, it is partly absorbed, partly transmitted, and partly reflected. The relationship between the reflected energy flow and the incident energy flow is closely linked to the physic-chemical properties of the object.

Thermal remote sensing systems measure the radiation emitted by surfaces in the thermal infrared wavelengths (5–15 µm). The measurement is indicative of the surface temperature of the observed objects. Optical remote sensing systems measure the solar radiation reflected from surfaces in the wavelengths in the visible (0.4–0.7 µm), near-infrared (0.7–1.1 µm) and short-infrared waves (1.1–2.5 µm). Reflectance is defined as the ratio between the intensity of solar radiation incident on a surface and the intensity of the radiation reflected by it. The curve that describes the reflectance as a function of the wavelength is called the spectral signature: this depends on the chemical-physical characteristics of the surface, as well as on the observation conditions [29].

The objects or different types of surfaces can therefore be identified by defining their spectral signature or the characteristic function that expresses the fraction of radiant flux (reflectance) reflected as the wavelength varies. By detecting and subsequently analyzing the spectral signature, it is possible to spatially map the presence of different types of objects [30]. Furthermore, it is possible to identify anomalies in the vegetation because its spectral signature presents typical characteristics which depend on the content and type of chlorophyll and on the foliar structure. Remote sensing, therefore, allows us to discriminate between the different plant species and thus monitor their vegetative cycle. As mentioned, different objects return different spectral signatures: for example, the spectral response of water is characterized by a marked absorption in the wavelengths greater than the visible and near-infrared. In this way, a hydrogeological basin is easily recognizable in a complex context [31]. From these examples, by measuring the energy reflected from the Earth’s surface at different wavelengths, it is possible to obtain the spectral signature of the various objects present in a spatial map and thus extract the nature of the surfaces that compose it.

2.2. Sensors for Remote Sensing

The ITRES CASI-1500 hyperspectral sensor (Compact Airborne Spectrographic Imager) (Figure 1) is an airborne sensor that can detect the reflected solar radiation of the area under investigation in the wavelength range between the visible and the near-infrared.

The CASI sensor is a passive system of the push broom type and measures reflectance (380–1050 nm). In fact, it uses a line of detectors arranged perpendicular to the direction of flight. During the flight, the image is collected one row at a time, with all pixels in a row measured simultaneously, receiving a stronger signal because it observes each pixel area for longer than with whisk broom technology. A disadvantage of these sensors is that they can have variable sensitivity, so if they are not perfectly calibrated, they can cause banding in the data. The sensor acquires images with a linear resolution of 1500 pixels, which can be discretized into 288 channels ranging from ultraviolet to near-infrared field. The sensor has a 40° angle of view which allows it to process images with ground resolution up to 50 cm with appropriate flight planning [32].

A measurement of the solar radiation reflected in many contiguous channels—intervals of the electromagnetic spectrum—allows the detailed reconstruction of the spectral signature of the observed surfaces. This type of measurement is called hyperspectral as opposed to multispectral measurements of satellite sensors which, by measuring the radiation in a limited number of channels spaced apart, do not allow for appreciation of the spectral signature of the surface materials on the ground. A sensor for the acquisition of hyperspectral images allows us to measure the spectral signature of the surface in each pixel of the image and, therefore, to estimate numerous chemical and physical properties. The spectral signature, or spectral reflectance curve, provides the measure of the ability of a given surface to reflect the incident energy at various wavelengths. The reflectance curve is a function not only of the surface material but also of the environmental conditions (period of the year, physical and chemical condition of the surface) and acquisition conditions (position of the Sun in the celestial sphere with respect to the surface and the sensor) [33]. The photosensitive unit and the optics of the CASI-1500 sensor can be configured according to two different scanning modes:

• Spectral mode: The spectral resolution is privileged over the spatial one when it is necessary to measure the detail of the spectral signature in the entire range from 380 to 1050 nm, ensuring a wide versatility of applications.
• Spatial mode: The spatial resolution is privileged over the spectral one when you intend to achieve better cartographic detail and, at the same time, optimize the spectral channels located at specific wavelengths for targeted technical applications.

The inherent wealth of hyperspectral data makes the scans carried out with the CASI-1500 a collection of information about the territory at the date of the flight. The CASI sensor is configured in such a way as to maximize the spectral channels that can be acquired according to the required geometric resolution and the integration times required for the detection of the radiant energy reflected from the ground. For this work’s requirements, all available spectral bands were selected [34].

2.3. Methodology for Remote Sensing Analysis

The scientific protocol for multisensorial aerial remote sensing conceived, tested, and verified by researchers [35], has planned the following investigation activities on flooded areas:

• Defining the area to be monitored;
• Setting the spectral channels of acquisition;
• Flight planning and execution;
• Pre-processing of data (

  Table 1. Software Specification.

  Image Processing	Radiometric Correction	Geometric Correction	GIS Mapping
  Envy 4.5	IDL script	IDL script	ArcGis 10.2.2

  );
• Processing of thematic representations and comparative and multi-temporal analysis of remote sensing data (Table 1).

Preliminarily, the numerical maps and official orthophotos, three-dimensional models and every certified geodetic reference of the territory to be flown over were collected and arranged in order to structure the geographical database through which to plan the voices, systematize the acquired data, order the multiple elaborations on homogeneous layers and critically analyze all the layers with vertical queries, according to a cyclical process of targeted analysis and integral verification of the data. The dynamic acquisition and the consequent software processing consist briefly of the concatenation of the following data groups:

• Photographic, hyperspectral scans;
• Aircraft attitude and flight position data;
• Data from the permanent GNSS stations on the ground.

Figure 2 shows a flow chart with an indication of the crucial phases of the survey methodology.

The product of the post-processing phase is a georeferenced, multi-layer model with a wavelength associated with each homogeneous layer of data. This model returns the energy reflected by the scanned object, detected by the sensor, and recorded in a single pixel of the image. The geometric dimensions of the pixel are a function of the flight parameters performed. It is a particular type of digital image characterized by such spectral complexity that, currently, with the technological means available, it is impossible to translate it into a syncretic representation that gives visible awareness of the “depth” of data recorded in it. Therefore, the selective representation in true colors, false colors or in shades of gray is preferred, depending on the spectral information which is to be highlighted based on the expected discretizations and possible or experimental thematic integrations. By focusing on hyperspectral acquisitions only, a specificity known to the scientific community should be noted, which finds innovative applications in the field of urban and environmental landscape surveys.

The vertical interrogation of the hyperspectral image on each pixel refers to the precise reading of the mediated and classified electromagnetic values in the respective scan channels of the sensor [36]. That is, it refers to the spectral signature of the material present on the ground or of the set of materials that fall within the pixel area, a function of the wavelength of the incident radiation at the instant of the aerial scan in comparable environmental conditions. Being able to isolate the spectral signature of a material means knowing its fingerprint, an element that makes it uniquely identifiable.

The comparative reading of the thematic representations thus drawn up—which can be related to thematic maps—takes place in the geographic database, structured in the first phase of the procedure. In this software platform, the data expressed in pixels are summarized in thematic tables, and the Digital Numbers recorded in each pixel are processed through the various layers (raster images). The comparative reading of remote sensing data with the maps and historicized orthophotos of the territory under investigation, or with the previous hyperspectral acquisitions according to a multitemporal remote sensing approach of the territory, is useful in complex and highly man-made scenarios. The reading of such a rich mass of data is also facilitated by projection and dynamic navigation on DTM (Digital Terrain Model) [37] or DSM (Digital Surface Model) [38] models, respectively, for natural or man-made scenarios drawn from archive repositories or products following photogrammetric or LIDAR instrumental acquisitions [39]. Also, in this case, multitemporal analysis is an essential reading key in integration with the previously described raster data. The discretization of the materials presents on the ground through hyperspectral remote sensing activities cannot disregard the extraction of the spectral signature characterizing the specific material under investigation. Once the sample material has been identified in parameterizable environmental conditions, it is necessary to process the specific spectral signature in the electromagnetic segments that can be scanned with aerial sensors with specific sampling and sampling.

The hyperspectral images acquired during the flight, accompanied by information relating to the flight plan, are archived according to a hierarchical model necessary to guarantee the subsequent phases. We then proceed to the correction of the radiometric and geometric errors and to the georeferencing of the images [40]. The data relating to the flight are obtained with the use of data from the GNSS satellite tables and from data extracted from the inertial platform and the onboard GNSS [41]. The data returned by the onboard GNSS, measured every second from take-off to landing, contain the position and attitude of the aircraft (roll, pitch, drift). In this way, the georeferenced path traced by the aircraft associated with the hyperspectral images of the scanned area is acquired.

Two software procedures are performed (Table 1):

• Radiometric correction;
• Geometric correction.

The remote sensing data are digital images (raster) which represent the territory detected by means of a pixel matrix. Each pixel is associated with two variables, x and y, which identify its position within the image, and a positive integer, called DN (Digital Number), which represents the average radiance, electronically measured, of the area on the ground covered by the single pixel. A remotely sensed image, therefore, is a matrix of numbers that translates the amount of energy emitted or reflected by the object under examination into numerical values. In order to perceive this numerical matrix as a raster image, it is necessary to perform an inverse process of transformation from digital to analog on the DNs that compose it.

The detected images contain radiometric and geometric distortions due to the acquisition system (platform and sensor), the signal propagation medium (atmosphere), the shooting angle and the effect of the Earth’s curvature. Each raw remote sensing image is therefore affected by a certain number of defects and errors, depending on its type and resolution, which would prevent its correct use in cartographic applications. With radiometric calibration, signal distortions due to both the influence of external conditions at the time of acquisition (atmosphere, lighting, topography) and sensor malfunction (sensor calibration) are eliminated [42].

Radiometric calibration consists of a series of procedures with the aim of correctly estimating the reflectivity of the object observed starting from the measured radiance. We work on the physical meaning of DNs to correctly estimate the reflectivity of the object in the study. The quantitative usefulness of the remotely sensed data is, in fact, maximized by calibrating these data to a reflection value of the surface on the ground: the DNs of a remotely sensed image cannot be considered representative of the actual conditions of the surface due to a variety of effects, such as atmospheric dimming variable, illumination geometry and sensor characteristics. This process is carried out through three different phases, which represent three levels of subsequent refinement, for each of which additional information is required:

• Sensor calibration: conversion from DN to sensor radiance
• Atmospheric calibration: conversion from radiance at the sensor to radiance at the surface
• Solar and topographic correction: conversion from radiance to reflectance at the surface.

Radiometric calibration transforms the images taken into spectral radiance units through an estimate of the brightness of the target: To do this, the spectral data of the ground and the data of the aerial image must match.

Geometric deformations are due to factors related to the geometric relationships of the sensor–platform–target system, such as:

• Relative movement of the Earth and the aircraft
• Characteristics of the acquisition system
• The curvature of the Earth and the presence of reliefs on it
• Platform position changes during image acquisition.

The geometric correction of the remotely sensed images, therefore, has the purpose of eliminating the spatial deformations introduced by the complex phenomena and their interactions to generate a new image with the scale and projection properties of a cartographic representation [43]. The geometric correction returns an image geometrically congruent with the reference, linking the position of the pixel in the image with its actual position.

Geometric distortions can be divided into two main types:

• Systematic, present in all images acquired by the same sensor, and due to the rotation of the Earth and the curvature of the Earth’s surface;
• Unsystematic, present to different degrees in the images acquired by the same sensor, and due to the movements and orbit of the satellite.

Corrections to eliminate systematic geometric distortions are normally made by applying appropriate formulas deriving from the mathematical modeling of their sources, while all other types of distortions must be corrected by means of the analysis of control points (Ground Control Points [GCPs]) appropriately distributed within the image. The GCPs are reference points that can be safely located on the ground and easily identifiable on a digital image, defined both as image coordinates and as geographical coordinates. The ground coordinate values of the GCPs are used to identify, through least squares regression, transformation functions with the aim of deforming the starting image and placing it correctly in the chosen reference system. The result is an empty matrix linked to the initial one by the transformation functions: It is precisely the inverse of these transformation functions that allow you to fill the correct matrix, determining the pixel values of the correct image based on those of the image of incorrect starting. The correction of geometric distortions is therefore obtained by georeferencing and subsequent straightening (orthorectification) of the image acquired by the sensor: the georeferencing process consists of the orientation of the image in each reference system. The orthorectification process removes the scale variations introduced in the images by the elevation differences of the surface using a digital terrain model (DTM).

The geometric correction thus returns an image interlaced with the position and attitude data of the aircraft. This is because each linear sequence of pixel-images orthogonal to the run line is orthographically projected onto the cartographic plane, thanks to the use of the software and the DTM model. The resulting product is an orthophoto, that is, a georeferenced and geometrically correct image that constitutes an accurate representation of the surface of Earth and can be suitably used in the GIS environment or within Geographic Information Systems.

2.4. Methods for the Semi-Automatic Perimeter of Flooded Areas

Occasional floods, during a flood event, leave clear traces of debris in urban and rural areas. The spectral signature of the debris can be automatically identified over large areas by means of similarity criteria.

The Spectral Angle Mapper (SAM) is an algorithm for the evaluation of the similarity between a measurement and a spectral signature taken as a reference [44]. It represents a rapid classification method based on the analysis of the similarity between the reference spectral signature, defined for each class and the spectral signatures of the individual pixels. These signatures are treated by the algorithm as vectors in a space of size equal to the number of bands. Since each class is made up of a sufficiently large number of pixels, the representative spectral signature of the class is obtained as the average of the totality of spectral signatures of that given class and the similarity between the spectral signature of a generic image pixel and that of the classes is expressed as the angular distance between two vectors in the n-dimensional space [45]. Figure 3 obviously refers to only two bands. The pixel is then attributed to the class with respect to which it has a smaller angular distance: If this distance always exceeds a set maximum threshold, the pixel is labeled as unclassified. The angle is then measured in radians, between zero and n/2, where n represents the number of image bands. Thresholds characterized by small angles express high similarity between analysis and reference signatures, while large angles generate stronger dissimilarity [46].

From a strictly mathematical point of view, an N-channel hyperspectral measurement is a vector in an N-dimensional space. The SAM calculates the angular distance existing between the measurement and a collection of predefined spectral signatures and attributes to the measurement the same thematic label as the closest reference signature. The advantage provided by SAM is that two similar spectral signatures but with different brightness, whether due to the presence of shadows or to the normal intrinsic variability of the observed surface, appear to have low or zero angular distance [47]. This reduces the chances of misclassification due to these factors, especially in urban areas, where they are more frequent [48,49].

2.5. Evaluation of the Identification Method

The validation of a method consists of a collection of experimental data, to demonstrate the adequacy and reliability of the procedure used. In other words, for the results provided by the analysis through the application of a specific procedure to be accredited and have general validity, the procedure used must satisfy specific requirements (validation parameters), which certify, in fact, the success of the methodology.

In order to establish whether the procedure described can effectively represent a valid tool for emergency management, it is necessary to carry out an adequate evaluation of the results obtained. The evaluation of the performance of the algorithm represents the greatest criticality of the whole procedure as it represents the phase in which the actual potential of the method is estimated. In this phase, it is necessary to establish whether the benefits obtained from the application of the methodology justify the costs incurred in the entire analysis procedure. It is necessary to understand how to interpret the results obtained and establish whether it is possible to trust the data obtained from the analysis. Each of these points is characterized by a specific articulation bringing various factors into play while returning the information necessary to establish whether it is worth adopting the methodology in question.

Previously, we have already highlighted how, in the emergency phases immediately following a flood, it is usually impossible or at least difficult to reach the critical areas by land. In such conditions, it is at least inappropriate to collect the necessary information in situ because if it were ever possible to reach the disaster areas, the emergency operations necessary for the safety of people could be hindered. For these reasons, it was considered appropriate to base the process of evaluating the results on photo interpretation by exploiting the tools available from the GIS platforms.

In photo interpretation, georeferenced information is collected according to pre-established objectives and criteria using remote sensing: Vast experience is required to recognize and geometrically define the elements that make up the area under observation. In the beginning, the information from the ancillary data is exploited by using zoom factors to get an overview of the area in question. Having acquired sufficient landscape-morphological and agronomic knowledge, we move on to a detailed examination of the various elements. The images subject to photo interpretation were collected using a Phase One iXA sensor, a medium format camera with an 80-megapixel resolution. The camera operates with a high dynamic range to ensure brilliant shots in different lighting conditions. It is equipped with fast sync lenses controlled by a processing unit for eliminating blur. The camera has been synchronized with the flight management system to ensure simultaneous hyperspectral scanning shots with the Itres CASI-1500 sensor. The high-resolution optical camera allows you to perform aerial shots of detail on the area flown over. Although not suitable to produce topographic cartography, the non-ortho-rectified nadiral images, carried out jointly with the hyperspectral scans, allow for integration with the high spectral resolution provided by sensors such as the CASI-1500 with a high level of detail in the visible. This allows you to better describe the phenomena observed with hyper-spectral data and resolve any ambiguities of interpretation. The Itres CASI-1500 sensor and the PhaseOne iXA camera have been configured for the best synchronic acquisition with respect to the reality of the places.

To make a comparison between the images taken by the two sensors (CASI-1500 and PhaseOne iXA), we worked in a GIS environment. The high-resolution images were compared with those returned by the SAM algorithm to obtain an estimate of the extent of the areas affected by debris deposits. Through the application of the polygonization technique, we then proceeded to estimate the areas identified and carry out a comparison with the calculation of the percentage estimation error.

3. Results and Discussion

3.1. Geomorphological Structure of the Area under Analysis

In the period from 11 to 12 August 2015, the Ionian coastal area in the north-east area of the Calabria Region in southern Italy was affected by exceptional rainfall, concentrated both in terms of quantity and intensity, causing particularly serious effects in the urban centers of Corigliano Calabro and Rossano (Figure 4).

The observed effects on the ground were considerable, particularly in tourist and hotel structures, infrastructures, economic and productive activities and goods, with significant inconvenience to the resident population and the numerous tourists who flock to these areas in the summer. Very serious damage was recorded to the road network and to the water and fuel networks. The area is part of a complex orogenic chain, consisting of tectonic units from the lower Cretaceous to the Paleocene with European vergence, which in turn overlap starting from the lower Miocene on the carbonate units of the Apennine chain in formation.

Morphologically, the entire area is made up of a series of subparallel ridges-oriented North-South, separated, at the major watercourses, by valley incisions. It can be divided into three altimetric bands with different geomorphological characteristics: hilly, median, and coastal [50]. The hilly area is made up of gneissic formations with associated granite bodies. It has a uniform morphology, deriving from the lithological characteristics and the state of alteration of the outcropping rocks. Where the alteration has changed the technical characteristics of the rocks, reducing them to inconsistent sands, the slopes are almost always steep; where the rocks retain their original characteristics, the slopes are steep to sub-vertical. The median area presents transgressive Miocene sequences with basal conglomerates passing through sandstone and arenaceous limestone.

The Pliocene-Pleistocene sequences of conglomerates and sands appear in succession, with the transition to gray-blue marly clays, which, towards the top, pass to sands and closing conglomerates [51].

This area has a non-uniform morphology due to the variety of outcropping lithotypes. Therefore, it is steeper in the areas where the cemented conglomerates emerge and sweeter and more undulating in the outcrop areas of the clays and finally, where there are strips of marine terraces sub-Quaternary plains whose morphologies are elongated parallel to the coastline and degrading towards the sea [52]. Finally, the coastal area slopes gently toward the sea and is made up of pebbles, gravel, and sand. The plain itself is furrowed by a dense network of artificial canals that convey the water and by the terminal rods of the waterways that deposit alluvial material.

The geomorphological structure of the hydrographic basins is very important as the hydrogeological instabilities in progress (active landslides and erosion phenomena) provide the solid material that the watercourses transfer to the sea on flood events and which, in turn, the sea distributes along the coast. The slopes of the valleys that subtend these waterways are often affected by widespread landslides. The type of movement of landslides is varied. In the upper part of the basins, where the rocks of the altered sub-layer mainly emerge, the phenomena of sliding and collapse prevail, mostly in the quiescent phase, while the transition belt, where the sand-so-conglomeratic formations emerge, is characterized due to the presence of erosive phenomena of the rill erosion type [53] and where clayey formations are present due to creep-like phenomena and areal erosion phenomena. The intense urbanization and anthropization, which have affected a large part of the coastal area over the years, have strongly interfered with the depositional processes and forms, which, at present, are mostly inactive and with numerous watercourses that are harnessed, channeled, sometimes ducted in their terminal sections with reductions in the outflow sections.

3.2. Flight Planning

On 14 August 2015, a flight for hyperspectral and photographic remote sensing was performed in the time slots before dawn and local solar noon (

Table 2. Flight specifications.

Air Carrier	Sensors Installed	Flight Data	Flight Data
Piaggio P166-DP1	Itres CASI-1500	Flight altitude = 2039 m	10 run lines
Speed = 222 Km/h	PhaseOne iXA	Cross overlap = 37%	Run line length = 22.9 km

). The Itres CASI-1500 sensor has been configured for the best synchronous acquisition with respect to the reality of places. The Itres CASI-1500 sensor has been configured for a 36-band spectral acquisition with a pixel resolution of 1 m on the ground. After establishing the same flight altitude and the percentage overlap value between the scan run lines, 10 parallel run lines were designed to remotely sense a territory of approximately 141 km². Given the characteristics of the sensor and the type of expected spectral data, the aerial scan was planned in the time band of the local solar zenith, that is, between 11:30 and 15:35 local time.

Given the timeliness of the activity, the georeferencing of remote sensing data was planned only with the use of the GPS antenna installed on board the aircraft. The dynamically acquired data were processed with data from the global navigation and observation databases of the GNSS constellations [54]. In order to complete the thermographic flight before the local sunrise, 20 out of 24 planned run lines were acquired, excluding those relating to the steepest and most high-altitude areas. For reasons of aircraft autonomy in relation to flight techniques, 9 out of 10 planned run lines for hyperspectral and photographic remote sensing were acquired, excluding the most extreme run line on the coastline.

3.3. Hyperspectral Images Analysis

In this study, the hyperspectral images detected were subjected to an initial exploratory analysis using an algorithm that applies a sampling in the shades of red in relation to the presence of vegetation for the representation of the vegetated areas (shades from red) (Figure 5).

The representation of hyperspectral data is returned in false color, associating the radiance measurements in the wavelengths of 808, 675 and 542 nm, respectively, to the red, green, and blue channels. This representation enhances the greater or lesser density of vegetation with shades of red and the different types of artificial coverings in blue [55]. As already mentioned, occasional floods during a flood event leave evident traces of debris in urban and rural areas. The spectral signature of the debris can be automatically identified over large areas by means of similarity criteria. This approach is particularly effective where the spectral contrast between the debris and the affected surface is high, such as in the case of roads, railways, and other artificial surfaces. In rural areas, where the spectral contrast is low, the visual interpretation of hyperspectral data in their representations in natural color and in false color allows for identifying the path of debris flows more effectively.

Figure 6 compares the data extracted from the Alluvial Management Plan representative of the hydrographic network of the area affected by the flood with the hyperspectral data to return a multi-layer reading of the territory to support the results deriving from the activities analysis of the same.

The hydrographic network is the complex of river collectors that collect the surface water outflows, together with the corresponding solid outflows and convey them to the terminal section of the basin. The hydrographic basin is that portion of territory that conveys the meteoric precipitation waters into the main river through the formation of a dense network of secondary tributaries. Rivers and streams, therefore, represent the terrestrial phase of the water cycle. In fact, they originate from that part of the rainfall, which, once it reaches the ground, remains there as surface runoff water following the most sloping lines of the territory to the sea. Each hydrographic basin is separated from the adjacent basins by the watershed line and ultimately encloses within it a hydrographic network formed by the main watercourse and the tributaries that feed it. The advantages of establishing a direct relationship between network length and rainfall can be manifold. The current flow is a spatially and temporally integrated output which in turn depends on the precipitation dynamics. Consequently, the observed discharge at the outlet of a given catchment reflects the way in which the previous precipitation inputs in the contributing area have been stored and routed through the different landscape units. On the other hand, adequate characterization of the length allows making a direct inference on the basin-scale effect of the expansion and contraction cycles experienced by the river network in response to unstable hydro-climatic conditions.

In Figure 7, the layers relating to the hyperspectral survey in false red shades are superimposed with the hydraulic risk map.

The hydraulic risk map has been extracted from the Hydrogeological Structure Plan [56], which defines the hydraulic and hydrogeological structure of the territory belonging to the regional hydrographic basins through the identification, perimeter, and classification of the areas with hydraulic and geological danger for the safety of the people, for functional damage to buildings and infrastructures, for the interruption of the functionality of socio-economic structures. The Plan also aims to promote soil maintenance and defense works as essential elements to ensure the progressive improvement of the safety conditions and environmental quality of the territory, as well as to promote actions and interventions necessary to favor the best hydraulic and environmental conditions of the hydrographic network, eliminating the obstacles to the flow of floods, the good hydrogeological and environmental conditions of the slopes, the full functionality of the defense works essential to hydraulic and hydrogeological safety.

From the analysis of Figure 7, it is possible to highlight that the hydraulic risk areas are concentrated in correspondence with the major watercourses and accompany them to the outlets into the sea. These areas are geometrically delimited by risk bands that extend to the edges of the watercourse, which identify buffer zones. To these are added the more densely populated areas in correspondence with the two urban centers of Corigliano Calabro and Rossano.

3.4. Spectral Angle Mapper (SAM) for Detecting Debris Deposits

Subsequently, it was decided to subject the detected images to classification, now calibrated, for which an excellent result was obtained in the radiometric correction phase. Since the first step of the classification consists in training the algorithm [57,58] by defining the Region of Interest (ROI), it was necessary to identify an RGB synthesis that could facilitate this step. The display of the image only in shades of gray would, in fact, have made the delimitation of the ROIs more difficult, and the risk of choosing pixels that are not representative of the classes would have increased [59]. For this reason, a false color synthesis has been identified that is able to accentuate the differences between the different areas very well.

For the SAM algorithm processing of the remote sensing data in this mission, the reference spectral signatures were collected in the same CASI images with a data-driven approach corresponding to the major waterways [60,61]. Figure 8 shows some examples of the results of the SAM processing, limited to the remote sensing area. The very high-resolution detail allows a visual quantification of the damage. The examples are obtained from the run line delimited by the coordinates [16°29′57.3168“ E 39°38′5.6333″ N], [16°30′24.2933“ E 39°38′51.6569″ N], [16°40′27.5992“ E 39°35′49.6822“ N], [16°40′0.3634″ E 39°35′4.4604“ N]) which appears to be representative of the heterogeneity of the entire overland area.

In Figure 8, we can evaluate (in yellow) the perimeter of the areas where the debris is deposited following the flood event. The largest areas are in correspondence with the urban centers already highlighted in the map of Figure 6 as areas with high flood risk. A first estimate of the surface affected by debris in this strip was calculated in the GIS environment based on the SAM algorithm [62,63,64], the visual interpretations of the false color and natural color representations of the CASI data.

In Figure 9, the layers related to hyperspectral data, hydraulic risk map and attention zones have been superimposed.

The areas of attention are areas in which there is information on instability with elements of hydraulic and geological criticality [65,66,67]. These are areas where the need to carry out specific investigations aimed at defining the problems has been indicated, and the level of danger has been determined. From the analysis of Figure 8, we can see that in the vicinity of the urban centers of the municipalities of Corigliano Calabro and Rossano, there are areas of attention. Comparing these areas with those bounded with debris deposits by the SAM algorithm (Figure 6), we can see that the areas are in proximity. This indicates that the flood event led to a large deposit of debris where hydraulic and geological criticalities had already been reported by the Hydrogeological Structure Plan. The flood phenomenon of 12 August lasted for a total of about 16 h, while the maximum intensity was reached after 6 h, between 02:00 and 8:00 [68].

The risk assessment guidelines available on infrastructures [69] reflect the need for considering quantitative procedures and multiple hazards. One of the main causes of bridge collapse around the world is related to hydraulic processes during floods and/or Geotechnical events such as riverbed scour at the bridge foundation or landslides [70]. The monitoring of the damages produced by natural events is therefore also fundamental for the assessment of extreme actions on structures and infrastructures and for the assessment of the resilience of cultural heritage, and for the robustness of infrastructures [71,72,73,74].

The data collected dynamically and available, cartographically and numerically, represent a central tool to support the policies of risk mitigation for the identification of intervention priorities, the ownership of the funds, and the planning of soil remedial measures.

In Europe (Italy–EU Partnership Agreement), the indicators of the population at risk of landslides, floods and coastal dynamics are used for the assessment of the effectiveness of the structural fund’s measures. For this reason, it is essential to develop an automated procedure for the quantitative reconnaissance of the territory and not only after a natural catastrophe.

3.5. Evaluation Results

The objective of the final phase of the study is to obtain an evaluation of the performance of the SAM algorithm in estimating the areas covered by debris. The areas affected by the deposit of debris identified through the automated procedure were analyzed in a GIS environment. In order to obtain an estimate of the extent of the zones, the high-resolution images were compared with those returned by the SAM algorithm (Figure 10).

The high-resolution image in Figure 10a allows easy identification of the areas covered by debris, and the comparison with Figure 10b allows us to evaluate the ability of the SAM algorithm to identify these areas. In order to obtain a numerical result from this evaluation process, the areas affected by the debris were appropriately bounded in a GIS environment. The physical identification of a real object takes place in a GIS system through a geometric primitive, which can be a point, a line, or a polygon. It is necessary to uniquely identify which of these primitives to use to effectively describe each geosite in its geometry. A point can be useful to indicate a position in space and can refer to different types of sites, but it does not provide information on their extension and does not allow any type of area processing. The same is true for lines that additionally only contain length but no width data. The optimal solution is to use polygons, but here too, there may be problems and the risk of losing topological information. In fact, trying to describe a site in its physical boundaries leads to the definition of a series of jagged polygons between which contiguity cannot be created. It is, therefore, necessary to identify standard types of polygons capable of accurately delineating each object present in the physical world but which can adapt to all the different combinations that occur in reality. The geometric primitive must represent the object in its entirety by organizing homogeneous objects in a single layer.

From the analysis of the different zones identified by the SAM algorithm and from the comparison between high-resolution images and the state possible in the GIS environment, evaluate the extension area of the zones through polygonization.

Table 3. Comparison between flood areas detected in nadiral photographic and in hyperspectral images using SAM detection.

ID Zone	Photographic View (m2)	SAM Detection (m2)	Error (%)
24,631	758	725	4.35
24,635	1098	995	9.38
24,639	1892	1811	4.28
24,642	2532	2316	8.53
24,651	1524	1435	5.83

compares the measurements of the perimeter polygons and the debris deposit areas identified in the high-resolution images and identified by the SAM algorithm for some sampled areas.

In Table 3, we calculated the percentage error between the estimate of the areas affected by the deposit of debris evaluated on the GIS platform through polygonization and those identified through the application of the SAM algorithm: We evaluated an error that falls within the range of 4–9%. This error difference is due to the state of wear of the road surface; in fact, it was verified that where the state of wear made the color shades of the debris congruent with those of the road surface, the identification error of the SAM algorithm increased.

4. Conclusions

This paper presents the operational protocol adopted for the representation, mapping and monitoring of the territory following the alluvial phenomenon that occurred in the areas of Corigliano Calabro and Rossano in Southern Italy. The use of aerial remote sensing was justified by the difficulties in field detection operations due to the obstacles represented by the debris deposited following the flood. Given the impossibility of terrestrial surveys, the research group opted for the use of the aerial multisensory platform consisting of an Itres CASI1500 hyperspectral chamber complete with an INS inertial platform and GPS antenna.

The maps extracted from national datasets for hydrogeological risk management were compared with maps obtained from remote sensing activities carried out in situ. The integration of the sensors with specific flight plans and the comparative reading of the respective data on specific software platforms has led to the discretization of the big data in anomalies automatically divided into polygons without the need for direct measurements and, in so doing, it represents an excellent tool for risk prevention.

The methodology has proved to be particularly effective for the identification of debris deposits on road infrastructures, however showing limitations in the evaluation of alluvial deposits in fields. In order to address this problem, more attention should be given to the topographic variable and to the selection of the optimal spectral angle, which will improve the classification accuracy. Furthermore, given that the similarity measure proved to be insensitive to gain factors, as the angle between two vectors is invariant with respect to the lengths of the vectors, this benefit can mitigate possible effects of different lighting found in the different areas explored. Finally, although the performance of the SAM-based model has been positively evaluated through a comparison with photo interpretation, the limitations found in this study suggest future research to discover the potential of the SAM algorithm to generate better prediction results through a different segmentation of the spectra, and through an enhancement of the SAM algorithm with the use of methodologies based on Deep Learning.

The methodology described in this work can be applied for the assessment of damage in severe flood events, in which it is at least complex to reach the areas affected by the event by land. In the future, the acquisition of more case studies in homogeneous geographical areas will allow the development of early warning procedures for the management of multi-risk.