Can very high resolution satellite imagery and artificial intelligence modernise wildlife surveys in Africa?

Introduction

The populations of African elephants and White and Black rhinos have fallen across Africa over the last 40 years. Wildlife surveys are essential for monitoring changes in population numbers and the movement of species across landscapes. They provide a valuable source of data to inform critical decisions on how best to conserve threatened species. Traditional survey methods in savannah landscapes use aerial counts completed by humans in a plane, but this can be costly and labour-intensive, with variable results.

In 2022, there was a growing interest in using <1m high-resolution satellite imagery and artificial intelligence (AI) technologies to survey large mammals (over 1m in size) across vast and hard-to-reach areas.

Connected Conservation Foundation led a cross-sector collaboration to explore whether these technologies can provide a practical and feasible method of reducing the time and cost of wildlife surveys in savannah environments. A summary of the study findings can be found in this blog, but you can download the full white paper here.

Figure 1: Elephants detected in Pleiades Neo imagery 

Approach 1: The human eye only, to find and classify different species based on field-based expertise and insight

Madikwe's conservation experts with substantial local field experience manually scanned the satellite imagery with the human eye and identified 110 large mammals across the 200 km2 Area of Interest (AOI). Conservation experts used their in-depth knowledge of the reserve and species behaviours in this landscape to record their understanding and classification of the sightings.

The below images were taken at a 1:100 scale from the 30 cm satellite image captured at the same time as the plotted sightings for the first satellite image above. Using field data, it was concluded that the bottom left image contains only rhinos, and the animals in the rest of the images are elephants. This example indicates how difficult it is to differentiate between rhinos and elephants in 30 cm imagery taken close to the nadir line (the point on the Earth’s surface directly below the satellite). 

Approach 2: Computer vision and open-source (pre-trained) neural network models, trained on a set of synthetic animal images taken from drones and down-sized

Dimension Data Engineers identified three state-of-the-art object detector models as being best suited for rhino classification: RetinaNet (using the ResNet50 backbone), EfficientDet (EfficientNet backbone) and YOLOv5 (using the CSPDarkNet and a PANet neck). 

For the purposes of rhino detection, no satellite training data was available for these endangered species. Object detection approaches require an incredibly large amount of training data to ensure that the model can reasonably learn differentiating features. Therefore, this approach developed a synthetic training data set. This involved cropping rhino images taken from varying resolution drone and aerial imagery into bounding boxes, then downsized to the resolution of the satellite sub-image in the training set, (10,000 training images across approaches).

When running these trained models on 30 cm satellite data of Madikwe Game Reserve, the accuracy of results for 30 cm was unsatisfactory due to the high number of false positives provided by the models. However, based on the spread of false positives compared to the synthetic ground truth, these off-the-shelf models could potentially provide a heatmap ‘where to look’ approach to potential detections. This may be a useful option for conservation managers. (see below on candidate for future exploration). 

Approach 3: A combination of computer vision and a bespoke convolutional neural network model, trained on a set of synthetic 3D animal images

For this approach, Dimension Data Engineers used open-sourced computer-generated imagery of the animals to create a synthetic data set to train a convolution neural network (CNN) model for animal classification. Samples are seen in 3D top-down, representing the outline and size of the animals, as seen from the nadir line. The images were downscaled to match the satellite imagery resolution and stitched into the satellite imagery.

Heterogeneous vs homogeneous environments

The complexity of dealing with heterogeneous environments opposes typical image processing methodologies. Complexity in an image is synonymous with ‘noise’ or random variation of brightness or colour, making it extremely difficult to isolate one region of interest from the background. 

Time of satellite pass-over

The orbit timings for the Pléiades Neo constellation of satellites were a major constraint for performing wildlife surveys in Kenya. Capture times for geographies nearer the equator fall around 11.30-12.30 pm local time in the heat of midday. African wildlife often ventures out into the open in the early morning to feed and drink, then retreat to the shade to avoid the heat of the midday sun. In our Kenyan study site, animals were often concealed by vegetation at the time of satellite acquisition. The project focus moved to the lower latitude in South Africa and Madikwe Game Reserve, where earlier tasking allowed for better conditions for wildlife sightings and approach testing.

Shadows

Shadows in remotely sensed imagery are produced by different objects, namely clouds, trees, boulders and mountains. These shadows have the potential to impact the accuracy of information extraction. For example, when elephants are bunched together, shadows and blobs can merge and field teams get different results from humans eyeballing the imagery, ground-truthed data and AI. In this 30 cm image, taken at a Dam in Madikwe, a group of elephants are under the trees with shadows. It is difficult to accurately count how many are in the yellow box. Human detection of this image identified five elephants, but this was debated by the field team, who sighted only four in the area. 

Additionally, when satellite images are captured at a higher incident angle, the shadows of the animal are more pronounced and play a more significant part in classification predictions. It is extremely challenging to provide enough synthetic training data for all shadow possibilities for each species. This project only focused on images captured close to the nadir line. Higher incident angles could be a future area of research.

Cloud cover 

When surveying such large areas, there is a high probability that some clouds will obscure important sightings, even on a largely cloud-free day.  

The Madikwe captures were tasked on optimal days, with excellent visibility and clear skies, with the image recording under 5% cloud cover. For a critical tasking, of the 57 sightings captured by on-the-ground teams during pass-over, one field-ranger sighting identified two rhinos at a known location. However, this sighting was concealed by the only cloud in the sky that morning. If this had been an annual wildlife survey, these two crucial sightings of a critically endangered species would have been missed.  

Results

The practical study found that, although some results showed promise, significant challenges remain in using satellite imagery to survey species in savannah situations. The algorithm detection and visual human inspection of large animals in 30 cm resolution satellite imagery (incident angle <15°), did not give conservation managers the required level of accuracy in individually counting and classifying species, in comparison to on-the-ground sightings.

Principle research findings included: 

Both human inspection of this imagery and AI model analysis could not discern between different species of large mammals in certain situations, including:

• When different species are interspersed together in the same area (elephants/rhinos, wildebeest/buffalo). For example, near watering holes.
• When juvenile species are mixing with other adult species. For example, the difference between a juvenile elephant and an adult buffalo or rhino.
• Additionally, when species such as elephants bunch close together, the shadows of several animals merge into one, resembling a ‘blob’, and it is difficult to count exact numbers.
• Cloud cover is a barrier to this survey method. During the study, on-the-ground field teams made many wildlife sightings at the approximate time of the satellite pass-over. However, even on a 96% cloud-free day, vital ground sightings of Black rhinos were obscured by the only small cloud in the region. This resulted in priority sightings not being identified in satellite imagery, which was unsatisfactory to conservation managers.
• Lastly for the AI, there was a lack of animal sightings in the imagery to function as labelled training data. Instead, novel approaches to creating synthetic training data were explored. 3D imagery, and downsized drone imagery, as a representative of real sightings, were used to train the algorithms, with limited success.

Ongoing research is needed to explore new species and situations where 30 cm satellite imagery can be of value. Specifically in homogenous landscapes or seascapes, where the objects of interest have a strong contrast to stand out from their background environment. Use cases could include monitoring colonies or herds of animals in hard-to-reach environments, where greater variability in counting accuracy can be tolerated and mixing between species groups is limited.

Only when aerial drone technologies enable image resolution beyond 30 cm to < 20 cm should these methods be pursued as a viable individual-wildlife-counting technique for mixed large mammals, in mixed (heterogeneous /homogeneous) savannah environments.

A potential use-case for AI monitoring of wildlife in satellite imagery

One notable outcome of this study is that rather than AI providing satisfactory individual counting and locations of animals, the techniques produced a heatmap style of potential sightings, acting as an indicator of ‘where to look’ for human-eyeballing validation within an image (see below). Scanning satellite imagery with the human eye is incredibly time-consuming and requires exceptional diligence, so consulting with an AI model to highlight areas of potential detection could be valuable. Field teams believe this may be useful for certain use cases, especially when locating (rather than individual animal counting) hard-to-find colonies or herds of certain species in remote areas.