Solar photovoltaic (PV) power generation is a vital renewable energy to achieve carbon neutrality. Previous studies which explored mapping PV using open satellite data mainly focus in remote areas. However, the complexity of land cover types can bring much difficulty in PV identification. This study investigated detecting PV in diverse landscapes using freely accessible remote sensing data, aiming to evaluate the transferability of PV detection between rural and urbanized coastal area. We developed a random forest-based PV classifier on Google Earth Engine in two provinces of China. Various features including Sentinel-2 reflectance, Sentinel-1 polarization, spectral indices and their corresponding textures were constructed. Thereafter, features with high permutation importance were retained. Three classification schemes with different training and test samples were, respectively, conducted. Finally, the VIIRS nighttime light data were utilized to refine the initial results. Manually collected samples and existing PV database were used to evaluate the accuracy of our method. The results revealed that the top three important features in detecting PV were the sum average texture of three bands (NDBI, VV, and VH). We found the classifier trained in highly urbanized coastal landscape with multiple PV types was more transferable (OA = 97.24%, kappa = 0.94), whereas the classifier trained in rural landscape with simple PV types was erroneous when applied vice versa (OA = 68.84%, kappa = 0.44). The highest accuracy was achieved when using training samples from both regions as expected (OA = 98.90%, kappa = 0.98). Our method recalled more than 94% PV in most existing databases. In particular, our method has a stronger detection ability of PV installed above water surface, which is often missing in existing PV databases. From this study, we found two main types of errors in mapping PV, including the bare rocks and mountain shadows in natural landscapes and the roofing polyethylene materials in urban settlements. In conclusion, the PV classifier trained in highly urbanized coastal landscapes with multiple PV types is more accurate than the classifier trained in rural landscapes. The VIIRS nighttime light data contribute greatly to remove PV detection errors caused by bare rocks and mountain shadows. The finding in our study can provide reference values for future large area PV monitoring.