Multi-scale assessment of wavelet-guided deep learning for reference evapotranspiration forecasting to support vineyard water management

Accurate forecasting of grass reference evapotranspiration (ETo), a climate-driven indicator of atmospheric evaporative demand, can support crop water requirement estimation, irrigation planning, water allocation, and agricultural water management. However, most previous studies have focused on ETo forecasting at single locations or over short lead times, with limited attention to how intrinsic ETo timescales can inform model design for multi-week forecasting or how forecast performance changes across regional and field scales. Here, we developed and tested a convolutional neural network-long short-term memory (CNN-LSTM) framework for multi-lead daily ETo forecasting in Madera County, California. Forecasted ETo was also coupled with empirically derived and forecasted actual crop coefficient (Kc act) to estimate actual crop evapotranspiration (ETa) using the crop coefficient approach. Continuous wavelet transform (CWT) analysis identified a dominant ETo timescale of approximately 93 days across the agricultural area, representing a stable subseasonal feature in regional ETo variability. Sensitivity experiments showed that this dominant timescale did not directly correspond to the optimal input sequence length (Lin); instead, Lin = 120 days provided the most stable performance and lowest overall error across the 56-day forecast horizon. These findings suggest that the CWT-derived dominant timescale can serve as a physically informed lower reference point for Lin selection, although the optimal value still requires sensitivity testing. At the regional scale, ETo forecast skill remained strong throughout the 56-day forecast horizon. At the vineyard scale, ETo extracted from the California Irrigation Management Information System (CIMIS) did not fully represent local atmospheric conditions, highlighting the need for site-level correction before field application. After correction, ETo forecasts remained highly skillful during the first 11 forecast days at both vineyard sites, with R2 > 0.90 and RMSE < 0.41 mm day-1, but declined substantially after approximately two weeks. As an extension beyond ETo, the coupled ETa framework showed that empirical Kc act forecasting was substantially weaker and more site dependent, thereby constraining ETa forecasting despite strong ETo forecasting performance. Overall, this study demonstrates that wavelet-guided deep learning can provide skillful multi-week ETo forecasts across regional and vineyard scales, offering a practical atmospheric-demand input for vineyard water management while identifying local ETo representation and Kc act uncertainty as the main constraints on crop-coefficient-based ETa forecasting.