Determining the population of an animal species for a given habitat (abundance) is usually a major challenge for humans and technology. Experts are very good at identifying animal species and counting or estimating their numbers but there are time and spatial constraints associated with this approach which limit continuous analysis of the dynamic development of populations.
Sensors, on the other hand, can be used multiple times and on an ongoing basis. Yet, they cannot provide the same level of information on an ad-hoc basis. The data they collect is initially erroneous, incomplete, and sometimes contradictory. Moreover, interpreting sensor readings is prone to ambiguity.
A detector reading can be
- An incorrect measurement,
- An incorrectly classified detection (classification in the wrong animal species),
- A correct measurement with correct classification (true positive, population: +1) or
- A correct measurement with correct classification of an individual that has already been sighted and counted (population: +0).
AMMOD is investigating probabilistic methods of sensor data fusion. The goal is to stochastically model various interpretations of the detections listed above. The expectation is to obtain a computed estimate of species abundance. This estimate is never perfect, but Bayesian statistical methods allow for the integration of background and contextual knowledge, thus enabling precise modeling of knowledge about populations and their habitats, and thus calculation of an expected value based on this knowledge.
The AMMOD stations are methodically based on algorithms for tracking objects in space and time, so-called object tracking. Here, the number of objects in the sensor's "field of view" and the "state" of each object (e.g. position and kinematics) are assessed. Since the spatial resolution of the sensors is not sufficient for a precise determination of the trajectories of single individuals in most cases, AMMOD uses stochastic motion models aimed at estimating populations.
The sensor model, on the other hand, describes the behavior of the sensor by stochastic means, which means that it includes values for detection probability or false alarm rate. In the interaction of the sensor and movement models, an estimated value of the population can be calculated iteratively, incorporating a variety of interpretations of the data as well as knowledge about the behavior of the species.
Geographic information systems (GIS) also make it possible to integrate additional contextual knowledge for sensor data fusion, since they provide information – for instance on built-up areas, vegetation, water resources or terrain characteristics that can be accessed in a georeferenced format. In view of the ambiguities in data interpretation described above, such background data can improve the validity of results. AMMOD will therefore also investigate to what extent these data can be formalized and integrated into the process of iterative estimation calculation.