Remote sensing applications have undergone major advancements based on the development of more efficient computational resources and RF technology. Tailored systems can be built with high acquisition rates at lower costs, facilitating the processing of larger amounts of information.

Commercial marine radars can be used to estimate surface patterns of water bodies and land entities [1,2]. However, the radars require intervention to obtain the complete clutter, which is usually not modified to display discrete targets.

Several systems that use radar-clutter have been reported. In [3], a sea-clutter retrieval and analysis was performed using a high-speed analog-to-digital converter (ADC) with a PCI interface to transfer and synchronize video signals. For example, the companies Miros and SeaDarQ offer sea-wave monitoring systems that can be used for oil-spill detection applications [4,5]. These systems have a similar hardware structure because they use standard radar outputs. The main inconvenience of this approach is the increased costs associated with robustness and wide-compatibility requirements, which are not necessary in every application.

In particular, this work considers the generation of navigation charts for rivers. The charts can be built using cartographic restitution via high-resolution radar/satellite images and high-precision GPS readings. This method offers charts with shoreline estimation with precision of approximately 1 m [6]. To achieve reliable results, it is necessary to apply techniques related to digital filtering on radar images available in the literature [7,8]. Some of the works employed image-processing techniques of rivers or beaches to estimate the corresponding shorelines; however, the system requires high-computational resources if the shorelines are to be updated continuously (i.e., as the vessel travels through the navigation path), and the cost of whole system (acquisition and processing) is still a concern [9, 10].

Another approach consists on the study of shorelines through high-resolution videos and RTK-GPS with accuracies of 1.1 m and 0.23 m, respectively [11]. In addition, the implementation of a Visual SLAM (Simultaneous Localization and Mapping) system has been considered [12]. Such a solution could be adapted to map river shorelines in real time by holding cameras on port, stern and bow. However, these systems have the disadvantage of video visibility and the associated limited detection range. The main justification for these methods and their expenses is the precision of the navigation charts. Thus, in the case of braided rivers, where the course changes frequently, the costs associated with the mentioned techniques cannot be afforded/justified because of the mandatory continuous chart updates.

Although the work reported in [2] used radar measurements, the novel approach is to employ the complete data, including the data that is discarded for the intended use of the radar, so that this whole clutter can be processed for the detection of river shores employing commercially marine radars. It is clear that well-known radar-image filtering techniques [7,8] cannot be directly used under this strategy. The system in [2] was complemented with a custom-made acquisition system that produced raw geographic information that properly transformed estimates the river coordinates. This prototype showed that it is possible to upgrade flexibility at reduced costs by sacrificing resolution compared to the use of bathymetries. The system must be optimized to achieve trade-off resolution, perform cluster definition, and minimize coordinate error with respect to the actual river shore. Safe navigation requires adequate signaling systems; in particular, for rivers with changing courses, traditional systems (e.g., buoys) are ineffective. As a result, alternative signaling strategies are required.

Thus, the motivation to propose cost-effective systems for navigation is two-fold. First, previous work [13,14] on river navigation assistance systems showed that shoreline mapping is possible, and further optimization is required to be considered as an effective navigation tool; and second, the Colombian government has indicated great interest in improving the use of sustainable fluvial transportation of cargo and passengers [15], as demonstrated by the large investments made recently. One of the most important challenges arises from the continuously changing nature of Colombian river courses and thus the changing nature of their navigable channels.

This paper discusses a system that uses commercial radars for effectively sensing shorelines. This work focuses on a more robust acquisition system and on alternative shoreline estimation using clutter optimization and an image-processing algorithm. First, both the hardware and the software of the system are described, and then the coordinate-transformation method that governs the acquisition and standardization of the sensed coordinates is presented. Next, the image-processing approach for generating continuous shorelines is described. Finally, the performance of the system is compared to official data to summarize system performance. Conclusions are drawn based on the performance evaluation.

The base level of the Magdalena River was selected for testing. Even though the shores at this point are relatively constant, a complete set of very accurate coordinates is available for comparison; therefore, the test will show if the complete estimation strategy produces adequate results. The information for shores is provided by Instituto de Estudios Hidráulicos y Ambientales (IDEHA) of Universidad del Norte, which conducts regularly scheduled bathymetries for the operation of the fluvial port of Barranquilla, Colombia [25]. IDEHA maintains both historic and updated data of the area.

First, the Furuno M1715 radar is located at a fixed point in the river, and the measurement is made. The acquisition system receives the echoes from the radar, and the relative distance from the radar is transformed into WSG84 coordinates using GPS and heading sensor (Section IIIB). Because the acquisition system captures and reconstructs the image from the clutter, the relationship between the coordinate (distance) and the pixel is obtained. The settings of the Furuno M1715 radar produce 140 pixels per 278 m. This relationship is important because the next step in the process is to translate the transformed coordinates into pixels to generate the necessary images for shore estimation. Fig. 5 shows a larger measurement area compared to Fig. 4 but using pixels as the measuring unit.

Next, a 2D Gaussian distribution is attached to each point of the cloud, as shown in Fig. 5. The sigma of the Gaussian is selected to connect as many close points as possible while keeping the estimation error to a minimum. As expected, only at the end of the process can the selected sigma be adjusted to produce a better shoreline estimation. This value can be reduced if the radar sweeps the target sector several times and/or the speed of the vessel that carries the radar is kept slow (for this work, the radar swept a target sector once traveling at 10 KPH).

After applying KDE, the probability density will be higher when more measurements of a point are aggregated. A heat map can therefore be generated that displays regions with increased confidence of the measurement. The heat map corresponding to the selected river section is shown in Fig. 6 (using a sigma equal to 2.5 pixels). From Fig. 6, it can be concluded that the “red regions” are those where the confidence of the measurement is higher, and hence the probability that the actual shore passes through those should be higher. To avoid clustering the images creating areas of decision, the goal of this work is to propose a method to generate a continuous shoreline that, in effect, provides full resolution. Thus, the idea is to connect those points of higher confidence while the line length within the image is kept to a minimum.

The technique is inspired by the Seam Carving algorithm for content-aware image resizing [24]. This algorithm establishes a number of seams (paths of least importance) in an image that allows reduction of an image by removing seams or extension of an image by inserting seams. By definition, a seam is either vertical or horizontal. In any case, an importance/energy function is evaluated per pixel by measuring its contrast with its neighbor pixels. Generating a seam consists of finding the path of minimum energy cost from one end of the image to the other.

Thus, the seam concept can be applied to Fig. 6 if the image is rotated and the heat map inverted. Image rotation is

required to “align” the more confident sections either vertically or horizontally, and inverting the heat map will transform the “red points” to valleys or points with less energy. Using matrix-based math software, such as iPython, rotation and inversion are handled as matrices facilitating the inverse operations. Fig. 7 shows the two seams (one per shore) once the image is returned to its initial position.

Finally, to evaluate the estimation using the seams, a set of official coordinates provided by IDEHA are transferred to Fig. 7 using the already defined coordinate-pixel transformation. Given the size of the image in pixels (900 × 1000), it can be concluded that the estimation using seams is adequate only by visual inspection.

Figure 5 Source: the authors.

Figure 6 Source: the authors.

To generate error measurements, the image is analyzed to generate the minimum distance (in pixels) from the seam to an IDEHA coordinate. Thus, the IDEHA coordinates are mapped into pixels by means of a matrix transformation. Because the IDEHA coordinates are a finite set of numbers, a histogram is generated using pixels (distance) as variables (Fig. 8). The use of iPython enables the use of the Python Data Analysis Library, better known as pandas. As described in its website, pandas is an open source, BSD-licensed library providing high-performance, easy-to-use data structures and data analysis tools for the Python programming language [26]. Pandas reports that most of the distances are grouped in the 0 to 4 pixel range, which validates the image-processing approach for shore estimation. Among other measurements (Table 1), the minimum distance obtained in this exercise corresponds to 1 pixel (1.98 m), whereas the maximum is equal to 11 pixels (22.2 m). However, for this system, the most suitable performance measurement corresponds to the Circular Error Probable (CEP). The CEP is a measure of the system precision and can be defined as the radius of a circle, centered about the mean, whose boundary is expected to include 50% of the radar measurement for a given coordinate. For the image-processing approach, pandas reports an ECP of approximately 6 m. Table 1 summarizes additional error measurements of the proposed system with respect to the officially accepted shoreline coordinates provided by IDEHA.

Finally, the seams are transformed into coordinates and plotted along the IDEHA points in a georeferenced image, such as Google Earth, to visually observe the result of the process (Fig. 9).

This paper presented a radar-based sensing system to estimate the coordinates for river shores. The proposed system is complemented with image-processing algorithms to generate continuous shorelines while avoiding resolution errors from clustering techniques. The proposed technique was shown to be technically viable and able to reduce the costs for updating navigation charts.

A mathematical model based on the projection of parametric curves describing the planet as an ellipsoid allows its use without any geographical restriction. Next, once the coordinates are expressed in the WSG84 standard, they are transferred to a 2D grid as image pixels. This translation allows handling the estimation task as an image-processing problem. The idea is to generate a continuous line that describes the desired shoreline. The system was validated in an actual environment using the data provided by IDEHA, which maintains accurate historic and updated shore coordinates for the base level of the Magdalena River. Based on the validation results, the system was found to perform with a resolution of approximately 1.98 m, which is equivalent to 1 pixel, and a Circular Error Probable (CEP) of 6.28 m.

Figure 7 Source: the authors

Figure 8 Source: the authors

Table 1

Source: the authors

Figure 9 Source: the authors

The use of the image-processing technique showed that the problem of shoreline estimation can be improved as more radar measurements of a defined sector are available. More measurements of a sector translate into a higher probability in the heat map and therefore force the seam to connect in that sector. Sweeping the radar several times and/or reducing the speed at which the radar moves in the vessel can achieve the goal of increasing the number of measurements. Improving the cloud of points (Fig. 6) allows reducing the sigma of each Gaussian distribution associated with a point, thereby reducing the wandering of a seam and reducing errors.

It is possible to improve the statistics with more accurate equipment; however, local cargo companies have accepted these indicators. The costs of the proposed technique are far more convenient than the use of satellite photos and/or bathymetries; however, to guarantee the output of the shoreline estimation mechanism, an initial costly method can be employed, so that error measurements can be defined, which later allows affordable navigation chart updating process using the proposed method.