The challenges of traditional quantitative irrigation methods cannot adapt to the dynamic actual soil moisture content and meteorological changes, and the existing methods based on soil moisture thresholds cannot fully solve the problems of hysteresis and adaptability, lack comprehensive consideration of meteorological factors and growth dynamics, and fail to consider the subtle sensitivity to soil moisture changes and processing efficiency limitations. To address the above challenges, we propose UFOGCN-

SPANet, a novel and computationally efficient architecture specifically designed for resource-constrained precision agriculture. Its core innovation lies in the cascaded integration of: (1) a linear-complexity Unit Force Operated Vision Transformer (UFOViT) that replaces quadratic self-attention with matrix associativity and cross-normalization for efficient global spatio-temporal feature extraction; (2) Graph Convolutional Networks (GCNs) for modeling spatial dependencies; and (3) a Salient Positions-based Attention Network (SPANet) employing a novel Significant Position Selection (SPS) algorithm to dynamically focus computation on the most informative contextual features, drastically reducing complexity while enhancing discriminative power. This unique combination directly addresses the critical challenges of computational efficiency and effective context modeling in real-world irrigation systems. Experimental results show that the proposed method outperforms traditional GNN models such as SAGEConv with 12 standard time series forecasting methods in key metrics, including accuracy, precision, recall, and F1-Score.