Macroscopic anomaly detection in paddy fields from UAV perspectives: a frequency-aware spatiotemporal mamba approach

Unmanned aerial vehicles (UAVs) provide high-resolution and temporally continuous observations, offering substantial value for wide-area paddy-field inspection. Timely detection of macroscopic anomalies such as fire, intrusion, lodging, and disease is important for field management and risk assessment. However, supervised methods rely on predefined categories and large numbers of annotated abnormal samples, which are difficult to collect comprehensively in field environments. Meanwhile, UAV ego-motion, viewpoint changes, repetitive crop textures, and multi-scale appearance variations further increase the difficulty of anomaly detection, limiting the direct application of conventional image-based and fixed-camera video methods to moving UAV videos. To address these challenges, this paper proposes TFMAD, a normal-only futureframe prediction framework for frame-level macroscopic anomaly detection in UAV paddy-field videos. TFMAD adopts a U-shaped encoder-decoder architecture, learns regular spatiotemporal patterns only from normal videos, and identifies anomalies according to the discrepancy between the predicted and observed frames. The framework contains three core modules: the Frequency Spatiotemporal Correlation Module (FSCM) models time-frequency variation and spatial correlation; the Parallel Spatiotemporal Mamba (PSTM) module models temporal variation and local structures through bidirectional pixel-level and block-level scanning; and the DifferenceGuided Fusion (DGF) module regulates skip connections according to the discrepancy between appearance and spatiotemporal features, reducing the direct transfer of abnormal information to the decoder. To balance high-resolution feature extraction and computational efficiency, the encoder further adopts Token Statistics Self-Attention (TSSA) with linear complexity. In addition, this study constructs URVAD, a UAV paddy-field video dataset for frame-level anomaly detection, and conducts experiments on this dataset. The results show that TFMAD achieves 95.83% Micro-AUC and 96.80% Macro-AUC, obtaining the best overall performance among the compared methods under the current dataset and experimental setting.