Counting elephants from space: An analytic framework for wildlife abundance surveys in heterogeneous terrain using satellites and AI

Estimates for species population size (abundance) are often obtained using aerial surveys, in particular for the African savannah elephant ( Loxodonta africana ). However, aerial surveys are costly and logistically demanding to implement, which often results in infrequent surveying (e.g. 4‐year intervals for elephants). There is growing interest in surveys using high‐resolution satellite imagery and AI detectors as an alternative, yet their operational feasibility and biases in real survey settings such as heterogeneous vegetated terrain remain largely unexplored. We develop an analytic framework for operational abundance surveys in heterogeneous vegetated terrain and apply this to a case study for elephants in Botswana. We illustrate the key stages of an operational survey, determining survey design constraints at a national scale; training an AI elephant detector (validation , test ); performing ground‐truth validation with the first synchronous satellite and aerial wildlife count in Botswana; and exploring bias‐correction modelling which explicitly accounts for canopy‐driven availability bias. Comparing the statistical properties of satellites with the existing national elephant aerial survey, we show that while satellites offer important operational advantages, including permit‐free imaging and quick spatial coverage, they do not yet provide a direct replacement for all aerial surveys. In particular, satellite abundance estimates are highly sensitive to corrections required for animals obscured by vegetation and have increased uncertainty associated with detections at 30 cm image resolution. Practical Implication : There is potential to use satellite surveys for identifying elephants in heterogeneous terrain. Within existing technology constraints, the current optimal strategy may be to use such data as a complementary tool, integrated with aerial surveys, rather than an immediate stand‐alone replacement. This is due to reliance on bias‐correction models and added uncertainty around detections at 30 cm satellite resolution. Three pragmatic recommendations for developing a satellite‐based methodology include: (1) developing an accepted bias‐correction method to address the issue of individuals obscured from the satellite by vegetation, using additional available data such as GPS collar data; (2) quantifying rates of species misclassification at 30 cm resolution; and (3) testing through simulation study whether integrating satellite and aerial survey estimates into an integrated modelling framework can enable earlier detection of changing population trends.