In Non-Line-of-Sight (NLOS) environments, reflected signals of human activity from different locations lose stable patterns, resulting in ambiguity in signal variation patterns, which poses a challenge for cross-spatial activity recognition. To address this ambiguity, we design a series of signal processing methods and proposed a cross-spatial signal classification model, overcoming the challenge of gesture signal variability in dynamic NLoS environments. Specifically, to eliminate environmental noise while preserving gesture features, we propose a gesture signal enhancement method that combines spatial features to improve signal resolution and distinguishability, better capturing dynamic variations. We also introduce a gesture signal segmentation method that uses signal fluctuation strength and temporal features to extract gesture segments from long signal sequences. Finally, a model that adaptively prioritizes critical features and learns spatial transformations is proposed to address the issue of classifying objects that may exhibit multiple forms within the same category. We conducted tests in three NLOS scenarios across multiple experimental environments, achieving over 98% accuracy in recognizing five gesture types, benefiting from LoRa’s long-range transmission capability, with a detection range of up to 45 m.