Abstract: The accurate extraction of non-agricultural habitats is crucial for building highly heterogeneous agricultural landscape, promoting crop yielding, and maintaining farmland biodiversity. Here, we constructed a new modified green-blue vegetation index (MGBVI) to extract the non-agricultural habitat information based on visible light image from an unmanned aerial vehicle (UAV), and compared it with seven other visible light vegetation indices. The three indices with highest accuracy were applied in two agricultural landscape sites dominated by divergent non-agricultural habitat composition to verify their accuracy. The results showed that MGBVI, normalized green-blue difference index (NGBDI), and green-blue ratio index (GBRI) had higher accuracy in extracting non-agricultural habitats, which had more advantages in distinguishing non-agricultural habitats and budding farmlands with the overall accuracy of 95.08%, 94.50% and 94.46%, respectively. The MGBVI had more accurate information recognition ability on the field boundary with sparse vegetation coverage in non-agricultural habitats. The MGBVI, NGBDI, and GBRI accurately extracted non-agricultural habitats in two validation sites, and the overall accuracy was higher than 94%. The accuracy of these three indices did not vary with different non-agricultural habitat types, indicating that the green-blue channel indices, such as MGBVI, showed higher availability and stability in the extraction of non-agricultural habitats from UAV images. Overall, our results provide technical reference for dynamic monitoring of non-agricultural habitats in agricultural landscape with complex topographical conditions.

Key words: non-agricultural habitat, agricultural landscape, vegetation index, visible light image, unmanned aerial vehicle