Dynamic graph learning framework based seasonal and trend decomposition approach for potato crop evapotranspiration prediction

Dynamic graph learning framework based seasonal and trend decomposition approach for potato crop evapotranspiration prediction

Efficient estimation of crop water requirements (ETc) is important for sustainable agricultural water management, particularly under increasing climate variability. Traditional methods lack a comprehensive analysis of dynamic patterns associated with crop evapotranspiration factors. To address these limitations, we propose a dynamic graph-based Dual-Graph Semantic Fusion (DG-DGSF) for ETc estimation. The multivariate time series is decomposed into trend and seasonal parts. This decomposition enables us to attain two dynamic graphs, Seasonal Dynamic Graph (SDG) and Trend Dynamic Graph (TDG), with their semantic characteristics extracted through Dual-Graph Semantic Fusion (DGSF). Each model is incorporated with the Dynamic Graph Learner (DGL) model and Graph Convolutional based on Recurrent Unit (GC-GRU) to analyse the trend and seasonal components. The DGL receives the trend or seasonal information to produce dynamic graphs, while GC-GRU combines the dynamic graph characteristics with the original series data. To effectively combine and extract the semantic characteristics from the trend and seasonal parts, a contrastive learning model is designed, followed by a supervised prediction model based on a multi-layer perceptron. The proposed DG-DGSF model was tested on data collected over two years (2023–2024) in Prince Edward Island, Canada. Three experimental locations were selected within the research farm: Location 1 consisted of loam, Location 2 featured sandy loam, and Location 3 contained loamy sand. The DG-DGSF model is compared with state-of-the-art models, including BiLSTM, GRU, GCN, BiGRU, LSTNet, DGDL, TPA-LSTM, and GCN-LSTM. The performance of the DG-DGSF is evaluated using numerous visual, statistical and probability metrics. The results demonstrated that the DG-DGSF model outperformed the benchmark models with the lowest forecasting error and highest ETc prediction rates, RMSE = 0.0469, MAPE = 0.120, NRMSE = 0.0431, KGE = 0.977, NSE = 0.963.