Abstract: Aiming at the bottlenecks of unstable endmember calibration and the lack of classification methods in fractional vegetation cover (FVC) retrieval from unmanned aerial vehicle (UAV) visible images in small-scale and complex land feature areas, this study took the Qinling ecological restoration area as a case study, and constructed a retrieval and classification method of FVC balancing accuracy and efficiency via the dimidiate pixel model. It quantified the systematic bias of FVC estimation induced by non-normal characteristics in the process of spatial aggregation. It also compared the robustness of three indices, namely visible-band difference vegetation index (VDVI), EXG (excess green index) and DEVI (difference enhanced vegetation index), under different heterogeneous backgrounds. The results showed that remote sensing estimation of FVC was scale-dependent. During image aggregation, endmember values converged as the data tended toward a normal distribution, resulting in a linearly increasing trend of FVC with the aggregation factor (R²>0.86) and a systematic increase of 0.07%-0.23% per doubling of aggregation. The overall classification accuracy of VDVI achieved 76%-80%, with a Kappa coefficient of 0.713-0.772; both were significantly better than those of EXG (54%-79%) and DEVI (65%-81%). This indicated that VDVI was the optimal general index for FVC retrieval and classification in small-scale and complex landscapes. The proposed method was suitable for artificial restoration areas with complex land cover types, providing a practical approach balancing accuracy, cost and efficiency for the application of UAV visible imagery in small and micro-scale ecological monitoring and refined land cover identification.