Historically, abundance and distribution data for a range of data species from large spatial areas were collected by surveying large distances along pre-defined tracks, or from visual aerial survey data from manned aircraft by trained volunteers. While this can provide reliable data which is simultaneously geo-referenced by the volunteers, it also provides relatively low spatial coverage over often vast spatial areas. More recently, to efficiently increase spatial coverage, data of this sort is collected using unmanned survey vehicles (or drones) which obtain high resolution images over vast spatial areas at relatively low cost. Drones can be programmed to work over very large spatial areas (outside the 'line of sight') and are becoming a routine alternative for surveying animals over large areas. One consequence of data acquisition in this way is that these high resolution images for very large areas must subsequently be scrutinised (by human eye) to identify the location and species of each animal in each and this processing can be prohibitively time consuming if undertaken by trained individuals. For this reason, an automated approach (or at least automated assistance) which more quickly enables processing these images is necessary if we are to routinely use drones to collect survey data and extract reliably geo-referenced data in a timely fashion. Regardless of the method of data collection used in these cases, very large sets of data which apply to a very large spatial area (>1 million km2) are returned. The subsequent modelling of this geo-referenced data therefore also often presents substantial computational challenges. As a result, compromises are often made between capturing genuine trends in the spatial patterns across the area which vary substantially in nature across the area of interest (e.g. allocating sufficient numbers of parameters to adequately capture realistic spatial patterns) and fitting a single model (or few models) for the area of interest. The former must be considered to ensure any surfaces are realistically complex and the latter is desirable in order to seamlessly predict across large areas. Spatially adaptive modelling is a recently developed approach to model surfaces with variable structure across an area (e.g. some surfaces need to be highly structured in some areas but are relatively flat in others) however choosing the flexibility of these models can be a computationally intensive even for more modest spatial areas. While to address this, the spatial area may (arbitrarily) be partitioned into segments for analysis, the challenge lies in `stitching' these parts together somehow post-analysis and decisions about any partitioning must be made in advance or iteratively as part or analysis. This process would be both time-consuming and ad-hoc. This project will involve two main parts: the development of an automated image processing approach appropriate for high resolution images and a `spatial tiling' approach permitting spatially adaptive modelling over large spatial areas. The data under investigation comprises high resolution multi-species data from the Namib desert which has been processed by human eye and will serve as a comparison for the automated image processing approach developed. These data will also be used to evaluate the effectiveness of the spatial tiling method at capturing spatial patterns in the data.