A residual LSTM Model with feature-agnostic training for reliable estimation of FAO-56 Penman-Monteith reference evapotranspiration for rice in data-sparse environments

Accurate estimation of reference evapotranspiration (ETo) is essential for irrigation management and water resource planning, yet the widely adopted FAO-56 Penman-Monteith (PM) method requires a complete set of meteorological inputs that are frequently unavailable in data-sparse agricultural regions. Although existing machine learning approaches for ETo estimation have performed well when trained on fixed feature sets, scenarios where one or more sensor inputs are absent are underexplored. This work proposes a Residual Long Short-Term Memory (ResLSTM) model trained with a random feature-masking strategy, enabling reliable estimation of FAO-56 PM ETo from any arbitrary subset of available meteorological inputs. The model is trained on 42 years of daily observations from four districts in the humid subtropical region of northwestern Bangladesh and evaluated across all 63 possible subsets of six input features, within the climatic context of this study region. The model achieves R 2 = 0.98 ± 0.0173 and R M S E = 0.1596 ± 0.0986 m m / d a y for the complete set of features, outperforming all baseline ML algorithms. To enhance transparency, SHAP analysis identifies solar radiation as the dominant predictor. The proposed approach provides a practical, sensor-fault-tolerant alternative to FAO-56 PM-based ETo estimation in resource-constrained agricultural settings.