AI-based animal monitoring for marine biodiversity conservation along the North Sea and Baltic Sea coasts

Coastal and offshore areas are highly relevant in the context of globalized economies and their demands for fisheries, transport and sustainable energy production. However, the ecological impacts of increasing human activity, such as noise disturbance and sediment dispersal from construction works and shipping traffic, could pose a threat to the biodiversity of marine ecosystems. By balancing marine food webs, controlling pests and dispersing seeds, marine birds are not only important for the conservation of biodiversity, but are also often seen as an early warning indicators of environmental change, as behavioural and physiological characteristics of bird populations are linked to changes in habitat quality. Spatial obervations of the distribution and size of bird populations are therefore needed to conserve biodiversity. Due to the vast extents and sometimes inaccessible nature of coastal and offshore areas, repeated airborne remote sensing surveys provide an efficient means of monitoring marine birds. However, the detection and classification of features on the ocean surface, such as animals, waves or man-made structures, remains challenging and is often achieved through time-consuming manual image inspection and annotation by trained experts.

Here, we present first results of an AI-based approach to automatically detect and identify different features and facilitate the monitoring of marine bird species and populations: Our study is based on approximately 2.5 million optical images with a ground resolution of 2 cm from 60 airborne surveys which were conducted by the German Federal Agency for Nature Conservation (BfN) along the German North Sea and Baltic Sea coasts between 2017 and 2021. Previously, images with bird sightings from some surveys have been annotated manually, enabling the training of a deep learning algorithm. Technical challenges for AI-based bird detection include a wide range of image exposure conditions, from low to high brightness contrast between objects and background, insufficient spatial resolution for relatively small species and tracking specific birds that appear in successive overlapping images to avoid double counting. Thus, our method uses a neural network approach (Faster R-CNN) to localise potential object candidates (e.g. bird) within an entire image, while a subsequent network classifier identifies the broad classification category of the detected object. In addition, spatio-temporal tracking of the detected features is included by estimating the most likely object displacement within successive images based on flight speed and camera motion along each survey transect. This workflow allows relatively efficient processing of large amounts of high-resolution imagery, as well as general classification of objects at an early processing stage.

Ultimately, our automated analysis workflow will contribute to the preservation management of biodiversity in the German North Sea and Baltic Sea by facilitating the repeated monitoring of bird populations.