It is difficult to distinguish between woodland information and water body information in mountain shadow, which seriously affects the accuracy of large-range woodland mapping. Making full use of remote sensing spectral features and not relying on DEM has become the main method to eliminate the influence of mountain shadow. This paper takes Zhangzhou City, Fujian Province as the study area, and extracts the woodland distribution of the whole city from 2013 to 2023 in combination with GF-1 WFV images. The study first extracts the enhanced vegetation features of the image of the study area by calculating the normalized difference mountain vegetation index (NDMVI) to improve the feature difference between the target and the shadow. Secondly, the NDMVI feature is introduced into the multi-scale stacked denoising autoencoder (MSDAE) to construct multi-feature parameters integrating spectrum, texture and remote sensing index, and carry out large-scale woodland extraction. The experimental results show that the identification accuracy can reach 91.34% at the actual verification points, and the average value difference is 3.84% compared with the forest coverage rate data published in Zhangzhou Statistical Yearbook. Among them, the woodland extraction effect of the image in December 2019 is the best, and the forest coverage rate differs by only 0.65%. The method proposed in this paper can effectively reduce the interference of mountain shadow on medium resolution woodland extraction, and improve the accuracy of woodland extraction.