FreqEdgeViT: A Scalable and Efficient Reliability-Aware Transformer for Large-Scale Agricultural Information Systems

With the exponential growth of data in modern agriculture, satellite image time series (SITS) has become an important data source for scalable information systems that analyze global crop distribution. However, processing these massive, high-dimensional data streams poses significant challenges; existing semantic segmentation models suffer from prohibitive computational overhead and lack scalability. Furthermore, they are vulnerable to high-frequency non-phenological perturbations and mixed-pixel boundary ambiguity, which degrades reliability in agricultural Internet of Things (IoT) applications. In this work, we propose FreqEdgeViT, an efficient, reliability-aware, and boundary-guided Vision Transformer. Designed for scalable SITS processing, FreqEdgeViT integrates a factorized spatiotemporal architecture with two novel mechanisms. First, in the temporal domain, we introduce a Phenology-Aware Frequency Filter (PAFF) combined with Reliability-Aware Token Merging (Ra-ToMe). This combination utilizes spectral analysis to filter environmental noise and dynamically prunes temporal redundancy based on signal reliability, significantly reducing data throughput requirements. Second, in the spatial domain, we propose a Boundary-Guided Spatial Encoder (BGSE) that enforces explicit geometric constraints to resolve edge blurring in mixed pixels. Experimental results on two public SITS datasets demonstrate that FreqEdgeViT achieves state-of-the-art accuracy with significantly reduced computational costs. The proposed architecture offers a scalable solution for processing large-scale agricultural data, providing precise support for crop policy formulation and enhancing the value of agricultural information systems.