Abstract: Ordos is a representative city in terms of desertification control and new energy development in China. Accurately capturing the spatiotemporal distribution information of photovoltaic (PV) power stations in Ordos City is of great significance for adjusting policies for new energy management. However, the phenomenon that different objects exhibit similar spectra poses a challenge to the feature extraction of PV power stations in a desert area. To address this issue, this study added pseudo-PV samples based on the distribution characteristics of land types in the study area and incorporated radar data based on optical data. Accordingly, multiple combinations of spectral, textural, topographic, and radar features were constructed. In combination with the random forest (RF), support vector machine (SVM), and classification and regression tree (CART) algorithms, this study assessed the feature extraction accuracy of different schemes on the Sentinel-2 and Landsat8 datasets. Finally, the features of PV power stations from 2016 to 2024 were extracted using the optimal scheme. The results indicate that the RF model combined with all four feature types demonstrated the highest accuracy on both Sentinel-2 and Landsat8 datasets. Furthermore, the inclusion of pseudo-PV samples and radar features can effectively reduce misclassification caused by the phenomenon that different objects exhibit similar spectra. National policy orientation determined the scale and spatial layout of PV power stations in Ordos City across different periods. Specifically, PV power stations in the city experienced a steady growth in area from 2016 to 2022 and rapid expansion from north to south after 2022, with an average annual growth rate 3.5 times that from 2016 to 2022. This study can provide technical support and data for optimizing the feature extraction methods for PV power stations and adjusting relevant management policies in desert areas.