The authors have declared that no competing interests exist.
Conceived and designed the experiments: PTF JF IS. Performed the experiments: PTF. Analyzed the data: PTF. Contributed reagents/materials/analysis tools: PTF JF IS. Wrote the paper: PTF JF IS.
We describe a method of identifying and counting whales using very high resolution satellite imagery through the example of southern right whales breeding in part of the Golfo Nuevo, Península Valdés in Argentina. Southern right whales have been extensively hunted over the last 300 years and although numbers have recovered from near extinction in the early 20th century, current populations are fragmented and are estimated at only a small fraction of pre-hunting total. Recent extreme right whale calf mortality events at Península Valdés, which constitutes the largest single population, have raised fresh concern for the future of the species. The WorldView2 satellite has a maximum 50 cm resolution and a water penetrating coastal band in the far-blue part of the spectrum that allows it to see deeper into the water column. Using an image covering 113 km2, we identified 55 probable whales and 23 other features that are possibly whales, with a further 13 objects that are only detected by the coastal band. Comparison of a number of classification techniques, to automatically detect whale-like objects, showed that a simple thresholding technique of the panchromatic and coastal band delivered the best results. This is the first successful study using satellite imagery to count whales; a pragmatic, transferable method using this rapidly advancing technology that has major implications for future surveys of cetacean populations.
“How many are there?” Is a question that is often difficult to address in ecology particularly for marine species that are generally inaccessible and cryptic. This is clearly demonstrated in whales where, despite their enormous size, robust population estimates are very difficult to obtain. The extreme size of whales means that they have a high per-capita rate of food consumption and hence a potentially massive impact on their prey populations as well as the marine ecosystem. Accurate population estimates are also essential to
Southern right whales have a circumpolar distribution in the Southern Hemisphere. The distribution in winter, at least for breeding animals, is concentrated in shallow coastal waters in the northern part of their range
Southern right whales were hunted extensively from the 17th through to the 20th century. The total number processed is conservatively estimated at about 155,000. The pre-whaling population was estimated at 55,000–70,000 dropping to a low of about 300 animals by the 1920s. After 1935 they were legally protected but over 3,000 more were thought to have been taken by illegal whaling in the 1960’s
Since the cessation of whaling several southern right whale breeding populations (Argentina/Brazil, South Africa, and Australia) have shown a strong recovery
Of current concern is the unprecedented mortality of southern right whales on their nursery grounds at Península Valdés, Argentina, in what are the most extreme mortality events ever observed in a baleen whale
The traditional methods by which cetacean population abundance estimates are obtained use counts of whales along transects from platforms such as aircraft or ships, or counts from land-based vantage points
A previous attempt to count whales using satellite remote sensing data and had limited success
Here we describe a method of identifying and counting southern right whales breeding in part of the Golfo Nuevo in Argentina using satellite imagery from the WorldView2 satellite count. This is an ideal location to evaluate our methods because every year, from July to November, whales concentrate in high densities to calve and mate. These enclosed bays are characterized by calm and shallow waters increasing the chances of obtaining images with optimum conditions of visibility.
We acquired a single WorldView2 satellite image of a region of the Golfo Nuevo Bay, the southern of two bays which separate Península Valdés from the mainland of Argentina (
The red box denotes the area of imagery acquired for this study. The grey area gives an indication of the possible swath width of a single satellite pass.
Golfo Nuevo, the southern gulf of the Península Valdés, is a roughly circular shaped bay and between 80 – 100 km wide. The sheltered waters attract southern right whales in great numbers and, together with a similar sized bay just to the north, they hold one of the world’s largest breeding aggregations of the species. This represents one of the best studied populations of southern right whales, with an ongoing programme detailing the natural history and ecology of the species
A section of a single WorldView2 image (Catalog ID: 103001001C8C0300) covering an area of 113 km2 and taken on the 19th of September 2012 was purchased from the commercial provider Digital globe. The image was chosen from the Digital Globe archive for three reasons:
It covers the middle of the Golfo Nuevo Bay, an area with a high density of southern right whales.
The timing corresponds with the middle of the breeding/calf rearing season, which lasts between July and November.
It is cloud free with a calm sea-state.
Sea surface waves have a very strong influence on the ability to detect submarine features
The image acquired consists of nine bands of information; eight colour bands with an on the ground resolution of ∼2 m per pixel, (Digital globe
We assessed the returns of each band over a cross section of pixels through whale-like features of two types; surface features and assumed submarine features (
Several of the images could be interpreted as whale pairs, or as a mother and calf, others may be displaying behaviour such as tail slapping, rolling or blowing. On several images there is a strong return at one end of the feature which is mostly likely the calluses on the whales head. Reprinted under a CC BY license with permission from British Antarctic Survey and DigitalGlobe.
The left hand figure is from a feature at the surface, the right hand figure shows a submerged feature. Note that while all bands show the surface feature, only band 5 (the Coastal Band) identifies the submerged feature.
Previous attempts to identify whales using IKONOS imagery show that attenuation of light through the atmosphere is weak in comparison to the two major components of image degradation; scattering from surface roughness of the sea and attenuation of light through the water column due to water turbidity
Using ENVI5 image processing software and ArcGIS automatic detection of whale-like features in the water column was tested using maximum likelihood supervised classification, unsupervised classification (isoData and k-means) and thresholding of specific bands.
Supervised classifications need the signatures input information of the pixel values for each class in order to classify the image. These signatures are usually manually input by the user. The algorithm then segregates all the pixels in the image into classes representing the signatures.
Unsupervised classifications classify the image into component parts based solely on information held within the image; isoData uses a clustering algorithm to determine the natural grouping of cells, while k-means calculates initial class means evenly distributed in the data space, then iteratively clusters the pixels into the nearest class using a minimum-distance technique.
Histogram thresholding
To construct a test dataset the image was divided up into a grid and whale-like features were manually digitized and coded into three classes: probable whale (features that were whale-shape and whale-sized) possible whale (including weaker signals, bubble slicks and some groups of seabirds are classed as possible whales
Visual inspection of the image showed that a number of offshore objects, that were both the right shape and size (5 – 15 m) to be whales, could be identified in both the colour and the panchromatic bands (see
Note that the higher resolution of the panchromatic band gives more detail, but it this increased detail also renders the object into several parts. Other bands show less detail, but have the advantage of homogenizing the object into one group of pixels, an important consideration when attempting to build automatic identification routines. Reprinted under a CC BY license with permission from British Antarctic Survey and DigitalGlobe.
Returns from the four automatic analysis routines were assessed against the manually digitized data (
Manually identified whales (top) have been broken into three classes; shapes that are whale-like and whale-sized are classed as probable whales, other objects are classed as possible whales, but may include bubble slicks and some groups of seabirds. The third class are objects identified only in the water penetrating coastal band, these are interpreted as sub-surface feature that are potentially whales. The bottom image shows the whale-like objects identified from the thresholding analysis of the coastal band.
Manually digitized | Unsupervied iso means | Unsupervised kmeans | Threshold Panchromatic | Threshold Band 5 | ||
total signals | 91 | total signals | 158 | 102 | 64 | 101 |
probable | 55 | probable matches | 44 | 42 | 43 | 49 |
possible | 23 | possible matches | 16 | 11 | 14 | 15 |
Band 5 only | 13 | band 5 matches | 1 | 0 | 0 | 13 |
total found | 61 | 53 | 57 | 77 | ||
% found | 67.0 | 58.2 | 62.6 | 84.6 | ||
% of probable | 80.0 | 76.4 | 78.2 | 89.1 | ||
total missed | 30 | 38 | 34 | 14 | ||
% missed | 33.0 | 41.8 | 37.4 | 15.4 | ||
false positives | 97 | 49 | 7 | 24 | ||
% false positives | 61.4 | 48.0 | 10.9 | 23.8 | ||
% good | 38.6 | 52.0 | 89.1 | 76.2 |
There are many objects in the image that resemble whales, but the question remains; how do we know it is a Southern right whale? The answer to this can be broken into three criteria used to identify any objects in remotely sensed imagery:
The object is the right size and shape to be a whale
The object is in a place we would expect to find whales
There are no (or few) other types of objects that could be misclassified as whales to cause errors of commission.
In this study we have digitized and automatically identified objects that are the right size (up to 16 m long) and shape. Although the size of the whales has an upper limit the lower limit is difficult to assess as the deeper the whale in the water column the less we are likely to see. The shape is generally ellipsoidal, although this can vary due to rolling, tail slapping and bubbles and other ripples associated with the animal. In the location of the study at the time the image was taken we expect to see a high density of whales in the image, especially mothers which, at this time of year, are forced to swim at the surface to support their calves.
There are only a limited number of other confounding artefacts that could cause errors of commission: No other large marine mammals are reported to frequent this bay, right whales are the only large whale species that regularly use the shallow calving grounds of Peninsula Valdés
On several potential whale objects there is a strong return at one end of the feature which is likely to be from calluses on the whale’s head, a feature which could aid automatic detection. Several objects identified as whales could be interpreted as pairs, or as a mother and calf, others may be displaying behaviour such as tail slapping, rolling or blowing (see
The top row shows examples of surface features that are probably bubbles from subsurface whales. Whether the whales are still under the bubble areas is difficult to ascertain. The lower row show clusters white dots, probably seabirds. Seabirds have been recorded to feed on whales at Península Valdés (see discussion). The third (and possible fourth) of these images shows a larger white object that could be a whale (or whale carcass) although once more it is impossible to tell with any certainty in this imagery. Reprinted under a CC BY license with permission from British Antarctic Survey and DigitalGlobe.
The results from the automated analysis suggest that a thresholding of the water penetrating Band5 returns the best results, finding 89% of features classed as probable whales in the manual count. Thresholding a single band is a very simple technique, although it does require some user input to identify the best thresholds. The greater accuracy of the technique (in relation to the more automated unsupervised analysis) needs to be balanced against the need for extra manual input in relation to other methods. These results however are promising and suggested that larger surveys over whole calving areas, which could potentially measure thousands of square kilometres, could be automated with a degree of success using these techniques.
The next challenge is to determine detection probabilities and understand whether counts from images can be used as a reliable index for population size, or presence. This paper shows that automated analysis of satellite imagery can achieve a good match with manual counts, but more work is needed to ensure that these manual counts are commensurate with the real number of surface whales. Once an estimate of visible whales has been formulated the ratio of visible whales versus invisible whales (those at depth or not at the breeding locality) is required to ascertain the total population size. One critical factor is estimating how deep the satellite sensor is seeing into the water column; the greater the penetration, the larger the proportion of the total population that will be identified. Penetration varies with water turbidity and surface roughness, two factors that may change over short time-spans and spatially within the image. Some estimation of turbidity may be made by comparison of the infra-red bands to the visible bands
The behaviour of right whales, with mothers calving in very shallow waters in protected bays, makes them an ideal candidate for the automated analysis of satellite imagery. The right whale population at Peninsula Valdés was previously thought to be recovering well, but recent years have seen persistent events of calf mass mortalities, suggesting major changes which require re-assessment; the latest available population estimates are over a decade old
We have shown that the use of current satellite imagery can be used to identify individual whales both at, and just below, the surface. The methods described here readily lend themselves to the calculation of population abundance estimates and suggest that behavioural patterns could also be elucidated. The automation of the methods means that counts can be carried out more quickly and efficiently than using traditional methods. This will allow a greater frequency of counts, both within and between years, that should lead to more robust population estimates, and the build up of a time series to asses trends. The important differences between our approach and a previous relatively unsuccessful attempt to identify whales from satellite imagery are the improvements in the on-the-ground resolution of panchromatic imagery and the use of the costal band (band 5) that penetrates to subsurface whales. These improvements allow a reasonable confidence to be assigned to the identification of individual whales thus allowing counts of whales in the wild as opposed to observations of animals in captive tanks.
A working system of whale population assessment by remote sensing will be an important new method that is potentially applicable other species of whale. Many species of whale breed in areas of calm water where, in order to protect their vulnerable calves, females remain close to the surface e.g. Humpbacks (
Our methods can potentially help providing within and between season population estimates, changes in distribution and use of the breeding grounds, both for right whales and other species of whale that breed in sheltered locations. Importantly, future satellite platforms planned in 2013 and 2014 will increase the on-the-ground resolution of panchromatic imagery from ∼50 cm to 34 cm and coastal band from ∼2 m to 1.24 m (Worldview3 planned launch 2014). This will result significantly higher quality imagery and therefore, greater confidence in identifying whales and differentiating mother calf pairs. Such improvements will also provide the opportunity to expand similar methodologies to other whale species.
This work forms part of Ecosystems programme and MAGIC within the Polar Science for Planet Earth (PSPE) strategic science framework of the British Antarctic Survey (BAS).