Drought stress prediction in Camellia oleifera seedlings using a deep learning hybrid model with temporal-spatial feature fusion

Abstract

Camellia oleifera, a distinctive and economically vital woody oil species in China, holds significant ecological and economic importance. However, the increasing frequency and intensity of drought events due to global climate change severely threaten its growth and yield stability. This study established controlled drought conditions in a greenhouse environment, and measured Soil and Plant Analysis Development (SPAD) values of two-year-old grafted container-grown seedlings to assess chlorophyll content and photosynthetic potential. Substrate moisture content (Volumetric Water Content, VWC, %), substrate temperature (℃) at upper, middle, and lower container positions, as well as greenhouse air temperature (℃) and relative humidity (RH, %), were monitored. A hybrid deep learning model, Temporal Convolutional Network-Bidirectional Long Short-Term Memory with dual attention mechanisms (TCN-BiLSTM-D2), was developed to predict SPAD values using these environmental variables. Results identified a critical substrate moisture threshold: plant mortality reached 100 % when VWC dropped below 5 %. Substrate temperature exhibited strong positive correlations with air temperature (r = 0.85–0.86) but negative correlations with relative humidity (r = -0.55 to −0.56), while substrate moisture exhibited strong negative correlations with both air temperature and substrate temperature (r = -0.82 to −0.67) and positive correlation with relative humidity (r = 0.30–0.37). SPAD values were significantly correlated with moisture in the middle and lower substrate layers (r = 0.16–0.63). Cultivars CL40 and CL53 exhibited significant negative SPAD responses to rising temperatures (r = -0.36 to −0.06). The model incorporated Feature Focus Attention (FFA) and Multiple Soft Attention (MSA), collectively termed D2, to dynamically weight input features based on their predictive relevance. This enhancement achieved exceptional performance, with a coefficient of determination (R²) of 0.982, Mean Squared Error (MSE) of 0.001, and Mean Absolute Percentage Error (MAPE) of 3.79 %. The TCN-BiLSTM-D2 model substantially outperformed conventional methods, including Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Recurrent Neural Network (RNN), and Temporal Convolutional Network (TCN). This framework enables non-destructive, high-throughput phenotypic monitoring and early warning of dynamic environmental stress, providing a robust tool for drought-resistance research in C. oleifera and practical support for the optimization of irrigation, the improvement of cultivation, and drought-tolerant breeding.