Refining classification of complex agroforestry mosaic landscape from drone‐based imagery: Implications for landscape management and conservation

Abstract Accurately and efficiently mapping landscape compositions is a conservation priority as natural habitats that are reservoirs of biodiversity are increasingly lost or modified. Globally, agroforestry is a rising sustainable land‐use type that benefits landowners and biodiversity, but these land‐use land‐cover (LULC) mosaics can be challenging to map due to the size and distribution of different LULC patches, requiring improved methodologies to map these areas and refine landscape‐based conservation management strategies. We classified an agroforestry landscape in West Java, Indonesia, which includes the habitat of a Critically Endangered primate, the Javan slow loris ( Nycticebus javanicus ). We used object‐based image analysis of high‐resolution drone (DJI Phantom 4 Pro, 2.90 cm/px mean resolution) and satellite (PlanetLabs, 3 m resolution) imagery, comparing how the resulting classifications differ and impact the classification of GPS waypoints collected during focal animal follows. Estimations of LULC types significantly differed between the classified drone and satellite imagery, affecting all habitat types. Compared to classifications from drone imagery, satellite imagery‐based classifications overestimated forest cover by an average of 18.81% and chayote (a vining crop) by 16.02%, while underestimating non‐chayote agriculture by 18.72%. Specific crop types were also classified from drone imagery that were impossible to classify using satellite imagery. Image classification differences are reflected in significantly different classifications of loris GPS waypoints, with an overestimation of the amount of waypoints in forest and chayote, and underestimation of waypoints in non‐chayote agriculture. Synthesis and applications . Together, these results show that significant differences in landscape classification between high‐resolution satellite and drone imagery can substantially change our understanding of slow loris habitat availability and use. Thus, it can be beneficial to use the highest resolution imagery to characterize species' habitats, whenever possible, particularly within mosaic landscapes. Additionally, image and map resolution should be explicitly reported. Less accurate area estimations of LULC types, particularly for forest and agriculture, directly impact conservation planning because species persistence may rely on these habitat patches and restoration of connectivity between them. Changes to forest and agriculture LULC can reflect the socio‐economic processes driving landscape changes.