Introduction

1The Paddle project (“Planning in a liquid world with tropical stakes”) defines one of the objectives of marine spatial planning (MSP) as “reconciling human uses of the sea with conservation”. A crucial starting point for this is to obtain information on these uses to inform future governance actions. One approach to this that partially overcomes possible data gaps on human activities is to extract spatially explicit information from satellite images available on Google Earth and process it using GIS (e.g. Quantum Geographic Information System software, QGIS). Currently, Google Earth is not widely used to inform MSP processes, yet it provides high-resolution optical satellite images (from Landsat) and aerial photographs that contain a wealth of information.

2This chapter proposes a standardised methodology (for the purposes of sharing and reproducibility) for extracting data from this rich source of information. In concrete terms, this involves superimposing a discretisation grid on a background layer of Google Earth images and then manually pointing and clicking on each element of interest (fishing boats, seaside resorts, sun parasols, etc.) within each pixel of the grid. Then, an automatic count of these features can be performed by QGIS. The features counted per unit of space and their combination can be considered as relevant surrogates for fishing and tourism activities, for example, allowing density maps to be produced. Our case study for this approach is located on the coastline of Pernambuco, a state in northeast Brazil (the Nordeste region) on the Atlantic Ocean. The main activities on this tropical coastline are tourism and fishing.

Materials and methods

3This section provides details for a standardised protocol for generating data based on a combination of Google Earth images and QGIS processing.

Tools

4Google Earth (GE) and QGIS¹ are two free platforms available on all operating systems, allowing any user to access optical satellite and aerial images, with a resolution on the order of a metre. This data can then be manipulated to extract relevant information: for our purposes, concerning the spatial distribution of anthropogenic uses.

Google Earth

5Google Earth is a software program that visualises the planet’s surface through a combination of aerial and satellite photographs. The satellite images cover the entire surface of the Earth and currently come from Landsat 8 (launched by the National Aeronautics and Space Administration, NASA, in 2013). The program selects coverage dates that minimise cloud cover and guarantee a minimum resolution of 15 m at any point on the planet. The maximum resolution of geographical locations depends on their interest. For example, aerial photographs of urban areas can be observed with a sufficiently high resolution to be able to distinguish each individual building, house and even car (resolution on the order of a metre). In this case study, we used GE images as a background layer in QGIS to count objects of interest. In addition, Google Street Views were used to corroborate the nature of the counted objects: a closer view often helped to differentiate between types of boats or infrastructure.

QGIS

6Quantum GIS (QGIS) is a widely used free GIS software. We used Madeira QGIS version 3.4 and installed the “Go2NextFeature3 2.00” extension. QGIS is a generic and user-friendly tool that overlays geographic layers and includes various useful features. It can be used, among other things, to view, browse, analyse, map, create, manage and export data. The relatively intuitive interface makes it easy to use even for a beginner, and the available extensions add even more functionality. The interest in using QGIS to collect and analyse data lies in its versatility (data collection and mapping), as well as in the fact that it allows working with different sources of information (satellite images, personal databases, data from agencies or institutes, etc.). In our case study, QGIS was a useful tool to conduct a count of features of interest in MSP: with the “Go2NextFeature” extension we were able to point and click on features in each pixel of the grid applied to the GE image backgrounds and thus semi-automate the count.

Method

7Our method combined QGIS software with the GE image layer to extract a database of human uses. The overlay of a grid layer allows the discretisation of the study area and then automatically counts categorised features of interest within this grid that can then be treated as spatially explicit indicators of anthropogenic uses of coastal spaces.

Discretisation grid

8The grid was constructed to cover the irregularly shaped coastline of Pernambuco (fig. 1), which extends approximately from longitudes 35.19° W to 34.79° W and latitudes 8.92° S to 7.39° S. The grid was composed of 29,295 cells of 220 m x 220 m, which corresponds to a coverage area of about 1400 km². The resolution was high enough to distinguish and count features of interest (boats, sun parasols, etc.).

Figure 1. Discretisation grid (in red) of the Pernambuco study area and a zoom on Recife (yellow rectangle)

9The discretisation grid was generated using the dedicated QGIS function in the “Vector” tab. The default grid produced by QGIS is rectangular, which was not suitable to cover the latitudinally extended Pernambuco coastline. To remedy this, we created and positioned pixel centroids using the “Generate points (pixel centroids)” function in the “Vector” tab. Then, the “Distance to the nearest hub (points)” algorithm included in the QGIS toolbox was used between the pixel centroids and the file containing the shoreline coordinates. This procedure resulted in the final discretisation grid by deleting the pixels located more than 2 km from the shoreline (Fig. 2).

Figure 2. Workflow for developing the discretisation grid

10Each pixel has an identification number and spatial coordinates. Once the grid is created and overlaid on the GE image background, it is possible to count the features of interest in each pixel and thus determine their spatial distribution. In practice, each pixel (identified by an ID number) is associated with a row in the QGIS attribute table, a spreadsheet used to retrieve the data. The QGIS “Go2NextFeature” extension allows the pixels to be listed (fig. 3) and the relevant information to be noted by point and click.

Figure 3. Go2NextFeature Extension Control Panel

Categorisation

11In this case study, we sought to identify objects that could be interpreted as indicators of human-induced pressures and that were identifiable in aerial views. In this area, where tourism and fishing activities are dominant, we counted the following objects: sun umbrellas, swimming pools, hotel infrastructure (four categories of hotel size: small, medium, large and very large), fishing gear (deployed nets and fish enclosures) and boats. Sun umbrellas, swimming pools and hotel infrastructure can be an indicator of tourism-induced anthropogenic pressure. Hotel infrastructure was subdivided into size categories to better describe the potential intensity of pressure created by tourism activity (fig. 4).

Figure 4. Size difference between two types of tourist infrastructure

12In terms of boats, a first category distinguished the use to which they are likely to be put (tourism or fishing) and a second category classified fishing boats according to their size: canoe-type fishing boat (canoa, non-motorised), raft-type fishing boat (jangada, possibly motorised, but low-powered and outboard), motorised fishing boat (inboard motor, with a deck or not), and finally boats for tourism. Each category of boat has a particular shape that can be recognised in aerial images (fig. 5). A canoa can be distinguished from a jangada by its size, the former being smaller. When there was doubt between two categories (between two types of engine, for example), the environment where the boat was located allowed us to determine its category. Motorised open boats are almost exclusively found in mangrove and estuary environments, whereas motorised boats with a deck are most often found in lagoons or at sea.

Figure 5. Illustration of the five categories of boats and the environments where they are most often found

Enumeration of the features

13Once the features to be counted are identified, it is time for the most laborious part of the job, which is to point and click on each feature under consideration in each unit of the grid. QGIS automatically counts the selected features and provides a tabular output. This point-and-click approach allows for visual transparency and traceability, which hopefully allows for future improvements and possible corrections.

14This method is made possible by two functions in QGIS. One allows the editing of several layers at the same time and the automation of an essential part of the count. This allows all layers to be put into edit mode at the same time (1 in fig. 6). To do this, the operator simply clicks on the “Add point feature” button (2 in fig. 6), selects the layer corresponding to the feature to be added and clicks on the feature in question (3 and 4 in fig. 6).

Figure 6. Screenshot of the counting area and application of the point-and-click method

15Another function is based on the “Count points in polygons” function, which is a Python program, offered in the “Vector” tab of QGIS. This automates the counting of points resulting from the point-and-click process (fig. 7), which would otherwise be a very time-consuming task. This procedure depends on the creation of a layer for each category of anthropogenic pressure, detailed in the previous section.

Figure 7. Control panel for the “Count points in polygons” functionality

16As soon as each pixel of the discretisation grid has been scanned, the “Count points in polygons” program can be launched for each object layer. QGIS then adds a new column to its attribute table with the number of points in each pixel. Finally, to centralise all counts, each count column of each feature layer is included in the attribute table of the count grid using the “Merge vector layers” functionality of QGIS (fig. 8).

Figure 8. Control panel for the “Merge vector layers” functionality

Indicators of anthropogenic pressures

17This section presents how to calculate indices of anthropogenic pressures (in our case study, related to fishing and tourism pressure along the Pernambuco coastline) from counting features in aerial images.

General calculation

18The main difficulty we encountered was the observable disparity between the count values obtained in each pixel for a given category and between the categories themselves. For example, in the “sun parasols” category, most of the values were less than 100 occurrences per pixel, while some reached 500 occurrences. The category “extra-large infrastructure” was far from this number of occurrences per pixel, representing only 48 objects in total. Such disparities prevent a normalisation with the maximum distribution value. We therefore decided to plot the distribution of the histogram of each (non-spatialised) count to allow for further examination. In doing so, we determined that it made more sense to normalise the results based on the 95% quantile of non-zero values in each category (fig. 9). We removed the null values because they are the most represented value in the count, but are not meaningful in terms of occurrences of the category. After normalisation, each type of object counted became comparable and could be included in a weighted sum. It was then necessary to find a suitable formula, taking into account the fact that each category does not contribute equally to the anthropogenic pressure on the environment. In order to represent the unbalanced effects, adjustable relative weights were assigned to each category, which can be expressed as follows:

19Using this calculation, we were able to provide an overall density map (fig. 10) of anthropogenic pressure. Its colours range from white (lowest density) to yellow, orange and red (highest density). The scale was calculated in relation to the pixel with the highest index value. In short, red represents maximum pressure, yellow/orange represents moderate pressure and white means that no pressure was identified.

Figure 9. Distributions of “sun parasol” histograms before (top) and after (bottom) the 95% quantile normalisation

Figure 10. An example of the conversion of raw count data (left) into coloured pixels based on the calculation of the tourism index (right)

Fishing activity

20A similar methodology can be applied to derive an index of fishing pressure based on counts of different types of boats (for an alternative source of data on boat monitoring, see Box 1). As an illustration, the following formula was applied successively to each pixel of the grid to calculate an overall index of fishing pressure per pixel:

these different categories, we chose a weighting that simply optimises the graphic representation of the overall index. However, this pragmatic choice could easily be modified by a different balance of weights if objective criteria of other kinds are identified by experts in the field.

Figure 11. Sample map showing the intensity of maritime traffic from AIS data in number of trajectories per pixel (2015, all types of vessels)

Tourism

21Similarly, we calculated an index of tourist pressure according to the following formula applied to each pixel of the Pernambuco coastline:

22Here again, the relative weights were chosen for illustrative purposes, to optimise the graphic representation of the overall index. This pragmatic choice can easily be modified by a different balance of weights if objective criteria of another nature are identified by experts in the field.

Results

23The algorithm used for the calculation of pressure indices aimed to produce two types of files: a file containing raw counts per pixel and maps of pressure indices related to fishing and tourism.

Counts

24The object tables for each count layer, as well as a summary layer, were exported in CSV format (comma-separated values, a text format with a comma or semicolon as separator), chosen because of its generic nature. This format facilitates sharing, storage and manipulation, both with Excel and with QGIS procedures. In addition, the procedure used to generate the anthropic pressure maps (in our case, related to fishing and tourism) can also create density maps of the raw count data. A total of 33,832 objects located along the Pernambuco coastline were counted and classified into the categories presented in Table 1.

Pressure index maps

25The resulting map of tourism-related pressures shows a linear and zonal distribution along the coastline in the south (fig. 12). The presence of many overlapping coloured pixels coincides with the location of the main cities of the state, such as its capital, Recife, or the seaside resort of Maracaípe. This indicates localised pressure, whose impacts are concentrated on a small area that can be identified by clusters of red or orange pixels surrounded by yellow pixels (fig. 12). The large white areas in the north overlap with those where mangroves are the dominant environment.

Figure 12. Extract of map resulting from the calculation of the tourism pressure index

26In the case of fishing, the trend is very different, indeed, the inverse. Fishing pressure is not concentrated around cities as in the case of tourism. Instead, yellow pixels are spread across all environments, from north to south (fig. 14). Although some impact zones are the same as those for tourism, the impact is much less significant. The pressure is more diffuse (fig. 13), i.e. exerted weakly to moderately on all environments, both in lagoon and mangrove ecosystems.

Figure 13. Extract of map resulting from the calculation of the fishing pressure index

Figure 14. Full maps of indices of fishing pressure (left) and tourism pressure (right)

Discussion

27This section discusses the limitations of the method, the key points to be retained and offers recommendations for possible future improvements.

Limitations of the approach

28One limitation to take into account is that during the enumeration process, problems of temporality were observed. The satellite images are selected by the program to minimise cloud cover; a selection that is necessary, but creates areas where two nearby images may correspond to different seasons. This can introduce a significant bias in object counts: for example, in our case beach umbrellas are highly dependent on the season. To remedy this bias in the count data, one could consider assigning an internal weighting factor that reflects, for each image, the season in which it was taken.

29Secondly, GE images record static objects, which are useful for assessing the spatial extent of infrastructure and features linked to human activities, but which only act as indicators of the presence of different activities that approximate the actual uses of marine areas. To deduce the use of areas from these indicators, it is necessary to formulate hypotheses based on other data: for example, what is the average range of a canoa, a jangada or different types of motorised craft? Which areas are most frequented by tourist boats and for which types of activity (walking, diving, recreational fishing, etc.)? What is the average occupancy rate of the different categories of hotels according to the season? What is the percentage of tourists who engage in activities at sea? These hypotheses can then be used to estimate, at a lower resolution than that of the counts, but still spatially explicit, the areas where the various marine uses are carried out. For example, the intensity of anthropogenic impacts on the seascape can be estimated by adopting an approach based on centres of gravity that combines the objects counted with the accessibility of different areas at sea, as proposed by Cinner et al.

30A further limitation is that in order to develop indicators of fishing or tourism pressures based on counting different objects, we had to “compare apples and oranges”. To compensate in part for the variable effects of the various objects on the occupation of space at sea, we proposed introducing weighting when integrating them in the form of an overall indicator of fishing or tourism activity. Clearly, this weighting is not insignificant in the final result, and it should therefore be the subject of consultation between experts and users of marine areas in order to best represent the reality on the ground. Ideally, a sensitivity analysis of the pressure maps produced with these weighting factors should be systematically carried out and opened up for discussion.

Points to keep in mind

31Our case study allows us to conclude that:

• The combination of QGIS and Google Earth can provide relevant information based on free data with global coverage, which is particularly valuable in a data-poor context.

• It is possible to create a spatially explicit database that is traceable, reproducible, easy to share and capable of feeding into MSP scenarios (and prospective impact assessments, see Box 2).

• The effectiveness of the approach is highly dependent on the resolution and temporality of the satellite images, as well as on the choice of objects to be counted, their respective weights, and the assumptions used to deduce the uses of marine areas.

Recommendations for future use of the approach

32The approach proposed in this chapter could be improved with one or more of these guidelines:

• Other databases could be used to describe anthropogenic pressure more precisely. For example, the Airbnb database could provide more detailed information on “tourism infrastructure” objects.

• New satellite images and improved image processing can provide better image resolution. For example, the Sentinel 1 satellite, using synthetic-aperture radar technology, offers the possibility of obtaining images, regardless of cloud cover, at a resolution of 10 m. This could limit temporal phase shifts between two neighbouring images.

• Deep learning artificial intelligence methods for image processing could be a relevant approach to automate the counting task. However, while this seems quite feasible for use at sea due to the uniform blue background, it seems more difficult for use on land.

• Maps obtained from images from other years could be compared to better understand the temporal variability of human activities.