Wireless Sensor Networks (WSNs) face critical data incompleteness challenges driven by hardware failures and energy constraints, which severely undermine environmental monitoring reliability. Although are frequently employed, Low-Rank Matrix Approximation (LRMA) methods often overlook nonlinear temporal dynamics and fail to discriminate structured noise from actual anomalies. This paper introduces the Adaptive Latent Feature Analysis with Fourier Embedding (ALFA-FE) framework, featuring two principal contributions: (1) dynamic Fourier embeddings that incorporate manifold-based frequency-domain regularization to flexibly capture multi-scale temporal patterns, and (2) a robust optimization scheme unifying Huber-norm loss with anomaly-sensitive constraints. Comprehensive evaluations across four real-world datasets reveal that ALFA-FE significantly outperforms seven state-of-the-art models in both reconstruction accuracy and robustness. By effectively balancing precise signal recovery with anomaly retention, ALFA-FE demonstrates strong potential for advancing environmental sensing reliability in resource-limited IoT deployments.