Reliable estimation of reference evapotranspiration (ET_o) is fundamental for irrigation scheduling, agricultural water management, and sustainable water allocation, particularly in regions with complex terrain and heterogeneous climate conditions. Although the Penman–Monteith equation provides accurate ET_o estimates, its application is often constrained by the requirement for complete meteorological inputs. To address this limitation, this study develops a hybrid deep learning framework integrating a Convolutional Neural Network (CNN), Temporal Position Attention (TPA), Bidirectional Long Short–Term Memory network (BiLSTM), and Adaptive Boosting (Adaboost), to enhance ET_o estimation accuracy from both temporal and spatial perspectives under limited data scenarios in Southwestern China. Daily meteorological observations from 95 stations covering three climatic zones during 1961–2020 were used to develop radiation–based (R_n), temperature–based (T), and humidity–based (RH) input configurations. Model performance was evaluated using both internal and external cross–validation schemes, with five non overlapping 12 years temporal folds used for internal cross–validation and spatial leave one station out validation applied within each climatic zone for external cross–validation, and compared with conventional empirical models and baseline deep learning approaches. Results indicated that the proposed TPA–CNN–BiLSTM–Adaboost framework consistently achieved the highest predictive accuracy across all climatic zones and validation schemes, with the coefficient of determination (R²), Nash–Sutcliffe efficiency (NSE), root mean square error (RMSE), relative RMSE (RRMSE), and mean absolute error (MAE) of 0.865–0.953, 0.853–0.950, 0.245–0.523 mm d^–1, 0.078–0.212, and 0.182–0.398 mm d^–1. The R_n based configuration achieved the best overall performance, especially in reducing error metrics compared with the T–based and RH–based configurations. Moreover, the proposed framework demonstrated smaller performance discrepancies between internal and external validation schemes compared with other models, indicating its superior robustness and generalization capability for both temporal extrapolation and spatial transferability across heterogeneous climatic regions. Overall, this study establishes an advanced BiLSTM–type ET_o estimation framework capable of accommodating diverse meteorological input scenarios in Southwestern China, providing effective technical support for agricultural water management and irrigation decision–making in climatically complex regions.