Assessment of riparian environments through semi-automated procedures for the computation of eco-morphological indicators: Preliminary results of the WEQUAL project

Within the WEQUAL project, a new method, called WEQUI, has been developed to evaluate the eco-morphological condition of GI. The method aligns with the EU Water Framework Directive recommendations (2000/60/CE) about classifying and monitoring the ecological quality of watercourses. WEQUI (WEQUAL Eco-morphological QUality Index) was inspired by several ecological, morphological, hydrological evaluation methods proposed in the literature and already adopted by local authorities for monitoring purposes (Lenat, 1988; Petersen, 1992; Ghetti, 1997; Braioni and Penna, 1998; Raven et al., 1998; Barbour et al., 1999; Davenport et al., 2004; Kleynhans et al., 2005; McGinnity et al., 2005; Rosegen, 2007; Siligardi et al., 2007; Kleynhans et al, 2008; Kemp and O’Hanley, 2010; Buffagnier al., 2013; Rinaldi et al., 2016). However, the WEQUI method differs from other methods, for the following reasons:

The purpose: WEQUI is intended to be applied for two main purposes:

classify the current eco-morphological condition of a watercourse, similar to what several other literature methods do, with special attention for water-courses where hydraulic structures are present. This kind of application is carried out recursively and allows to monitor the eco-morphological condition of a watercourse. In addition, in the case of design of a new hydraulic structure, one-shot evaluation of the WEQUI index allows the designer to become aware about the current environmental state of the site where the structures are planned to be built.

Basically, WEQUI consists of fifteen environmental quality indicators. For each indicator, five levels of quality are identified and characterized with short statements. Each statement is associated with a score. Once all the indicators are compiled, the final score is computed and classified, according to a proper rule. At the end, an evaluation of the eco-morphological quality of the watercourse is obtained. In the case of applications of type a), it accounts for the current quality condition, while in case of applications b), it accounts for the quality state expected some decades after the construction of new hydraulic structures. The fifteen indicators of the WEQUI index are assigned according to both qualitative and quantitative criteria. Quantitative criteria can be applied through semi-automatic procedures, while the qualitative criteria require a direct survey in the field or a computational system (Table 1). The number refers to the sequence of questions reported in WEQUI index. For each indicator, the score is assigned by choosing the most relevant answer in a list of five possible options. Depending on the indicator, the answer may require the observation of either the riverbanks, the channel or the whole reach corridor. Each answer is associated with an exponential score on basis two, from a minimum of one to a maximum of sixteen. Through this approach, it is possible to clearly identify the elements of low or high naturalness of the monitored environment. Adding up all the results obtained for each criterion included in the WEQUI index, it is possible to conduct an overall assessment of the eco-morphological quality on both to natural and to artificial rivers, characterized by engineering solutions (both traditional and soil-bioengineering solutions).

During the year 2017, according to the sensors, three flights have been planned on the municipality of Lagundo (Province of Bolzano, North-East of Italy) in order to monitor a portion of about 400 m of length of the Adige river for a total area of approximately 6.5 ha. The in-field surveys were performed using remote sensing technologies. All the collected data was then processed through semiautomated procedures described below.

Thanks to their high operative flexibility and the rapidity with which missions could be set up and executed, Unmanned Aircrafts Vehicles (UAVs) have been used to carry out the transportation of the sensors necessary for the monitoring operations (Ristorto et al., 2017). The use of UAV aircrafts for monitoring purposes is possible even when the access is difficult, which is very common in riparian areas (for instance due to slopes, high density stands or slippery surfaces). During the monitoring operations, the UAVs have been equipped with three sensors:

RGB camera Sony RX-100

The Sony RX-100 is a conventional camera with a resolution of 20.1 MP. During the photogrammetric surveys, the camera has been set in autofocus mode and with a fixed focal length at 8.8 mm, with a CMOS sensor of 13.2 × 8.8 mm, in order to achieve the largest angle of view. Meanwhile, the Micasense RedEdge-M is a camera able to acquire simultaneously images from 5 bands: Red, Green, Blue, Nir and Red edge wavelengths. The resolution of each image is 1 MP, with a fixed focal length at 5.4 mm and a sensor size of 4.8 × 3.6 mm. In every acquired image, the respective metadata geoinformation, which speed up the procedures of orthorectification, are also reported. Indeed, the Micasense RedEdge-M has the possibility to be connected to a GPS module. Depending on the flight altitude, the RGB camera and the multispectral sensor record images with different ground sampling distance. The minimum and maximum resolutions for the Sony RX-100 are 1.4 cm/pixel and 3.4 cm/pixel and Micasense RedEdge-M 4.1 cm/pixel and 10.2 cm/pixel, respectively, for 50 and 150 m of flight altitude. To obtain a useful dataset and orthophotos without distortions, the flight plan of the UAV (grid width) and the trigger signal to the sensor must be set to have a sidelap and an overlap of respectively 65% for the RGB camera and 75% for the multispectral camera. The missions during which we use these sensors were setup with parallel flight routes covering the portion of river targeted for monitoring including surrounding land to reach the overlapping percentages mentioned above, needed for this survey. The collected image datasets are processed with Pix4D software, obtain a mosaicking of the images. To orthorectify the final aerial images, during the flights, several ground control points have been measured using an RTK-GNSS device with an accuracy lower than 5 cm (Geomax Zenith 35 Pro). Beside these two sensors, a Yellowscan Surveyor LiDAR has been used to acquire a 3D model of the monitored surface. All the acquisitions are automatically orthorectified by the system thanks to the global navigation satellite module and the inertial measurement unit embedded in the sensor. The acquired data sets are also corrected using the RINEX files collected by the RTK-GNSS station placed in the middle of the monitored area (Figure 1c). Since the LiDAR is characterized by having a 2 echoes laser technique, from the obtained output, it is possible to generate Digital Terrain Model (DTM), Digital Surface Model (DSM) and Canopy Height Model (CHM) point clouds. The minimum resolution of the sensor is equal to 100 point/m². In order to increase the accuracy of the collected point clouds, post-processing kinematic correction are carried out in POSPac MMS 8 software. The flight mission with the use of this sensor foreseen, after a period of sensor calibration at the beginning of each mission, only three passages over the whole monitored area. Two at 40 m height over the two river side banks, one at 30 m height following the river’s axle. In this way, the system is able to acquire information related to the river, the vegetation and terrain profile of the riverbanks and of the lands nearby the river with very high details.

Figure 1

Due to the different weight of the sensors, the aerial fleet is composed of three UAVs with different propulsion systems, and flight performances (e.g., payloads and cruise speeds) (Figure 1):

Mavtech AGRI-1900

The AGRI-1900 is a fixed-wing UAV developed to accomplish missions that require high flight endurance; it ensures a flight time of 30–45 minutes/mission. The declared cruise speed is around 12–15 m/s, which permits to cover up to 50 hectares for mission in the best flight conditions. Nevertheless, this aircraft requires a wide surface without obstacles for take-off and landing operations. Usually, this UAV flies at 100–150 m of altitude, above ground level. Therefore, the use of this vehicle ensures the monitoring of large portion of land, but the collected data is affected by low resolution. The use of this aircraft has been planned to be used only in case of large monitoring. The payload of this UAV permits to install on-board only one sensor, like a camera, for a mission, which is attached at the lower part of the drone’s fuselage.

Beside the fixed-wing solution, a light and a heavy rotary-wing UAVs have been developed. The two rotary-wing UAVs differ from each other because of the different transportable payload. As such, the Q4E and the DJI S900 have 0.6 and 3.5 kg of payload, respectively. Nevertheless, the physical and mechanic characteristics of these UAVs influence the overall mission’s performances (Table 2).

Due to their characteristics, this solution of drones is more flexible than the previous one. They are capable to operate in extreme environments where wide surfaces for the take-off and landing maneuvers are not present. The rotary-wing drones have the capability to take-off and land vertical. Due to the small wingspan, these UAVs are limited to flights at an altitude lower than 70 m, covering less surface but with very high resolution. Besides, this family of drones has constraints on endurances, typically lower than 20 min and on cruise speeds, which can ensure a maximum coverage area of 5 hectares. Due to their payload, the Q4E can be equipped with two camera sensors, while the DJI S900 can be equipped with the LiDAR sensor. On the entire fleet of UAVs, an autopilot system is onboard. This tool automatically increases the aircraft stability and helps the pilot in controlling the vehicle during the mission. All the missions have been planned using the Mission Planner AutoPilot software, which helps draw the flight route as well as monitor all the flight parameters in real-time during the mission.

The collected data are then processed by a complex system called Inference Engine (IE). The IE can be considered as a machine learning application able to interpret and translate raw data into management information (Mazzetto et al., 2012). The IE is a set of algorithms and routines able to process, interpret and translate the raw data into information. Indeed, in literature, several solutions have already been proposed to carry out automatic crop and operative monitoring procedures in agricultural and forestry operative contexts (Gallo et al., 2013, 2017, 2018; Mazzetto et al., 2017). These procedures, integrated with a management information system, can be used to automatize all the elaboration and assessment procedures, from data collection to data consulting (Sørensen et al., 2011; Fountas et al., 2015a, 2015b). The final purpose of the Web platform would be the same as the management information system for agro-forestry contexts: develop an ICT tool that allows the end-user to obtain information in relation to the environmental quality of the monitored riparian area. Generally, all the procedures that characterize the IE are developed ad hoc according to the domain of monitoring in order to extrapolating all the information needed to compile WEQUI index. For this scope, some of the automatic procedures have been implemented taking inspiration from methodologies already described in the literature (Cavalli et al., 2008; Michez et al., 2013; Tompalski et al., 2017) and adapted to our case. While others have been developed specifically for this task. Currently, the development and implementation phases of the entire set of algorithms and procedures which compose the IE are still under implementation and testing. In fact, so far, only two algorithms have been developed: the classification of the land use (criteria 1, Table 1) and the assessment of the riparian vegetation along the riverbanks (criteria 11, Table 1).

In this preliminary phase, the algorithms have been implemented entirely in MATLAB. Then they will be translated into open-source coding language that can be easily integrated with web platform.

The two algorithms have been validated through visual comparison and manual measurements performed directly on the field. Indeed, to validate the algorithm for land use coverage, a visual recognition of the surface covered by vegetation, water and soil have been done using Arc-GIS software. This assessment has been carried out for the entire surface interested by the monitoring. The outputs coming by the visual evaluation, performed on the orthophoto collected by RGB camera, has been then compared with the results obtained by the semi-automatic procedure. For the validation of the algorithm for the assessment of the riparian vegetation along the riverbanks, five squared plots with an area of 225 square meters have been randomly selected and analyzed. In each plot, the entire aboveground biomass (bushes and trees) has been measured by two technicians. The surveyors, using a clinometer vertex laser and a GNSS-RTK unit, have measured the heights and taken the coordinates of each individual inside a plot. In this way, a shapefile has been set-up, where all the individuals have been classified as shrub or tree if they were smaller or bigger than five meters height. The above-mentioned equipment has been selected as the most appropriate for this scope, since they are commonly used in forestry survey and forestry inventories. In this case, for each plot, the validation has been conducted overlapping the output achieved from the automatic assessment (a georeferenced raster map) and the data collected by the field survey. The presence of correspondences between these two informative layers have been analyzed. When the collected points overlapped with an area with the same classification, the identification has been considered appropriated.

For each survey, the orthophoto obtained by processing the multiband acquisition is processed by the algorithm to obtain the percentage of land use coverage, divided into vegetation, water and soil. The results are then summarized in a thematic map of the land use. Figure 2 shows the flowchart of the developed algorithm.

Figure 2

A simplified Machine Learning approach is used to recognize thresholds usable to distinguish vegetation from water or soil defining several Regions of Interest (ROI) of the land use. Firstly, the user identifies and draws one or more polygons over the multiband orthophoto, identifying one or more vegetated areas to be used as a reference. According to the reflectance data collected through the five images (one for each band), the algorithm computes several Vegetation Indexes (VI), such as NDVI, NDRE, SAVI and OSAVI. Then, analyzing the statistical distribution of the computed VIs, the algorithm selects the most representative index through an automatic procedure that considers the distance between the two distributions (Figure 3). The 5^th percentile of the relevant index distribution is set as the lower VEGetation THReshold parameter (VEG_THR) to discriminate the area covered by vegetation (VEGETATION_ROI) within the whole study area. All the pixels with the selected VI higher than the threshold VEG_THR are classified as vegetation.

Figure 3

Then, the user draws one or more polygons to delimit one or more areas covered by water. Also, in this case, the statistic distribution of the VI considered as representative is evaluated within the selected sample areas. The 95^th percentile of the distribution is set as the higher WATer THReshold (WAT_THR) suitable to discriminate the area covered by water. All the pixels with VI lower than WAT_THR are classified as water within the whole study area.

In the end, the remaining pixels of the study area are classified as soil (SOIL_ROI). Figure 4 shows the graphical result of the classification of the land use algorithm, where vegetation is represented by green pixels, water by blue pixels and soil by brown pixels. For each class, the land use algorithm calculates the coverage extent.

Figure 4

The second algorithm developed allows to obtain the shape of foot and head of the riverbanks in several sections along the river axis, as well as soil and vegetation classification within the riverbanks. Figure 5 shows the analytical procedure of the implemented algorithm. In this case, the Digital Terrain Model (DTM) and the Digital Surface Model (DSM) obtained from LiDAR point cloud are used to calculate the Canopy Height Model (CHM), which is mandatory to be able to proceed with the other algorithm’s steps.

Figure 5

Firstly, the user draws the river axis over the RGB orthophoto, moving from upstream to downstream (Figure 6). Starting from the first point of the axis, the algorithm defines some transverse sections that are homogeneously spaced (Figure 7). The space between two transverse sections can be setup by the operator according to the accuracy required.

Figure 6

Figure 7

For each transversal section identified by the semi-automatic procedure, a moving average of the terrain elevation and its slope are computed (Figure 8). In this step, the algorithm searches the Points of Interest (POIs) necessary to delimit the riverbanks. The first two POIs identified (POI1 and POI2) represent the bottom and the top of the right riverbank respectively. The following POIs (POI3 and POI4) represent the foot and the head of the left riverbank. These points are identified analyzing if the first derivate of the terrain profile overpasses a well-defined threshold. The bottom of the riverbank is detected when the terrain slope exceeds the threshold, while the head of the riverbank is identified when the terrain slope drops below the threshold. The threshold can be modified by the user, depending on the slope of the transverse profile, in order to obtain the correct result.

Figure 8

The land use map obtained from the previous algorithm and the CHM calculated from the point cloud are then overlapped to the map where riverbanks are identified. The masking procedure permits to extract information related to the coverage of riverbanks and to classify the vegetation according to Kaufmann et al. (1999) as:

Herbaceous: if CHM < 0.5 m

As in the previous analysis, in this case also, the result is shown in a georeferenced thematic map, on which the classification of the riparian and floodplain vegetation are summarized (Figure 9), together with a table where the percentage of soil and classified vegetation of the riverbanks are reported.

Figure 9

All the results obtained by the semi-automatic procedure have been validated with those collected during in-field surveys. Due to the preliminary stage of the research, the validation of the results of the land use classification algorithm has been done only for two criteria considering the entire survey over Adige river. The visual interpretation consisted of overlapping the information obtained during the monitoring and the interpretation of the ROI during the postprocessing phase. Using ArcGIS software, through visual classification, the entire monitoring area have been classified as vegetation, water or ground. Then, the coverage area has been calculated for each land use class. The obtained results have finally been compared with the same results obtained through the automatic method. Table 3 shows the results. The results in Table 3 highlight that the implemented algorithm for the automatic land use classification is well able to recognize those areas where vegetation is present, while the identification of water and ground areas is affected by some uncertainties. Realistically, this result is related to possible inaccuracy in classifying zones of transition between water and soil and/or zones submerged occasionally. According to the hour and the season during which the UAV flights were carried out, portions of the collected images may be affected by a shaded band generated by the riparian vegetation. This shading effect causes a slight variation in the reflectance of transitional bands, affecting the data-processing and causing an overestimation of the ground content. A further cause of overestimation of the soil class may be caused by a similar reflection between thin water layers within the river, and because their reflectance is very similar to those of gravel, cobbles and boulders, which partially or scarcely submerged. As a consequence of this similarity, the algorithm identified the selected ROI with an accuracy of approximately 90 %.

For the validation of the algorithm used for the assessment of the riparian vegetation along the riverbanks, 38 trees and 37 shrubs (according to the adopted classification) have been measured and georeferenced during the field survey. The collected data has been used to develop a shapefile that was overlapped to the georeferenced raster map, output of the implemented algorithms. The semi-automatic procedure has been able to properly identify 82% of the individuals in the plots. This validation has been affected by uncertainty mainly in the classification of shrubs or trees close to the threshold (five meters) and errors due to wrong georeferencing. The discrepancy could be due to the intrinsic errors of the adopted measurement methodology or to the validation procedure. Indeed, the riparian environment can be considered as a very chaotic environment mainly if aboveground biomass is characterized by coppice or shrubs (brambles) that vegetate on very steep land. This steepness does not permit a proper positioning of the GNSS device and a proper distance between the surveyor and the tree when measuring the height, using the clinometer. Besides this, the presence of the coverage, mainly if the plant has leaves, affects the accuracy of the measurement causing a slight drift of the acquisition from the real position. In the end, also the inclinometer can be affected by errors because the height measurement is based on the distance and angles assessment between surveyor and tree. Therefore, a not correct identification of the top or the bottom of the plant can influence the measurement of the angles, so the height, up to a couple of meters. This error is emphasized when the assessment is performed at narrow distances. Indeed, it sometimes happens that the surveyor during the assessment was unable to reach the best place to perform the acquisition because it was not accessible due the presence of branches with thorns or for terrain steepness.