To address the large variation in target scales and cross-image semantic shifts in few-shot semantic segmentation of remote sensing images, this study proposes a multi-layer feature prototype network (MLPNet) to improve segmentation performance and generalization under few-shot conditions. MLPNet leverages multi-layer CNN features to generate diverse prototypes and incorporates an adaptive fusion module to dynamically adjust the weights of prototypes from different layers, thus enhancing the expressive ability of the prototypes. A semantic alignment module is designed to model the semantic relationships between support and query images through an iterative interaction strategy, effectively mitigating cross-image semantic shift. Furthermore, a class-specific region mining module is introduced to extract potential target information from low-confidence regions while suppressing erroneous activations in high-confidence regions, improving prediction consistency and completeness and enabling fine-grained segmentation. Compared to current mainstream few-shot semantic segmentation models, MLPNet achieves competitive performance under both 1-shot and 5-shot settings on the iSAID and LoveDA datasets, with the mean intersection over union (mIoU) on the iSAID dataset improved by 4.78% and 4.41%, respectively, compared to the baseline model. At the same time, stronger target capture capability and more robust segmentation performance are demonstrated on targets with complex categories and severe background interference. MLPNet effectively exploits the expressive ability of multi-layer feature prototypes and, combined with a cross-image semantic alignment mechanism, enhances the segmentation capability for complex targets under limited support samples through the collaboration of multiple modules, providing a generalizable and high-accuracy solution for few-shot semantic segmentation of remote sensing images.