Search

Soil moisture is a key parameter for the circulation of water and energy exchange between surface and the atmosphere, playing an important role in hydrology, agriculture, and meteorology. Traditional methods for monitoring soil moisture suffer from spatial discontinuity, time-consuming processes, and high costs. Remote sensing technology enables the non-destructive and efficient retrieval of land information, allowing rapid soil moisture monitoring to schedule crop irrigation and evaluate the irrigation efficiency. Satellite data with different resolutions provide different observation scales. Evaluating the accuracy of estimating soil moisture based on open and free satellite data, as well as exploring the comprehensiveness and adaptability of different satellites for soil moisture temporal and spatial observations, are important research contents of current soil moisture monitoring. The study utilized three types of satellite data, namely GF-1, Landsat-8, and GF-4, with respective temporal and spatial resolutions of 16 m (every 4 days), 30 m (every 16 days), and 50 m (daily). The gray relational analysis (GRA) was employed to identify vegetation indices that selected sensitivity to soil moisture at varying depths (3 cm, 10 cm, and 20 cm). Then, this study employed random forest (RF), Extra Tree (ETr), and linear regression (LR) algorithms to estimate soil moisture at different depths with optical satellite data sources. The results showed that the accuracy of soil moisture estimation was different at different growth stages. The model accuracy exhibited an upward trend during the middle and late growth stages, coinciding with higher vegetation coverage; however, it demonstrated a decline in accuracy during the early and late growth stages due to either the absence or limited presence of vegetation. Among the three satellite images, the vegetation indices derived from GF-1 exhibited were more sensitive to vegetation characteristics and demonstrated superior soil moisture estimation accuracy (with R2 ranging 0.129–0.928, RMSE ranging 0.017–0.078), followed by Landsat-8 (with R2 ranging 0.117–0.862, RMSE ranging 0.017–0.088). The soil moisture estimation accuracy of GF-4 was the worst (with R2 ranging 0.070–0.921, RMSE ranging 0.020–0.140). Thus, GF-1 is suitable for vegetated areas. In addition, the ETr model outperformed the other models in both accuracy and stability (ETr model: R2 ranging from 0.117 to 0.928, RMSE ranging from 0.021 to 0.091; RF model: R2 ranging from 0.225 to 0.926, RMSE ranging from 0.019 to 0.085; LR model: R2 ranging from 0.048 to 0.733, RMSE ranging from 0.030 to 0.144). Utilizing GF-1 is recommended to construct the ETr model for assessing soil moisture variations in the farming land of northern China. Therefore, in cases where there are limited ground sample data, it is advisable to utilize high-spatiotemporal-resolution remote sensing data, along with machine learning algorithms such as ETr and RF, which are suitable for small samples, for soil moisture estimation. [ABSTRACT FROM AUTHOR]

Accurate estimation of root-zone soil moisture (SM) is of great significance for accurate irrigation management. This study was purposed to identify planted-by-planted mapping of root-zone SM on three critical fruit growth periods based on UAV multispectral images using three machine learning (ML) algorithms in a kiwifruit orchard in Shaanxi, China. Several spectral variables were selected based on variable importance (VIP) rankings, including reflectance Ri at wavelengths 560, 668, 740, and 842 nm. Results indicated that the VIP method effectively reduced 42 vegetation indexes (VIs) to less than 7 with an evaluation accuracy of root-zone SM models. Compared with deep root-zone SM models (SM40 and SM60), shallow root-zone SM models (SM10, SM20, and SM30) have better performance (R2 from 0.65 to 0.82, RRMSE from 0.02 to 0.03, MAE from 0.20 to 0.54) in the three fruit growth stages. Among three ML algorithms, random forest models were recommended for simulating kiwi root-zone SM during the critical fruit growth period. Overall, the proposed planted-by-planted root-zone SM estimation approach can be considered a great tool to upgrade the toolbox of the growers in site-specific field management for the high spatiotemporal resolution of SM maps. [ABSTRACT FROM AUTHOR]

Accurate and timely prediction of soil moisture in orchards is crucial for making informed irrigation decisions at a regional scale. Conventional methods for monitoring soil moisture are often limited by high cost and disruption of soil structure, etc. However, unmanned aerial vehicle (UAV) remote sensing, with high spatial and temporal resolutions, offers an effective alternative for monitoring regional soil moisture. In this study, multi-modal UAV remote sensing data, including RGB, thermal infrared (TIR), and multi-spectral (Mul) data, were acquired in citrus orchards. The correlations between different sensor data and soil moisture were analyzed to construct seven input combinations. Convolutional neural network (CNN), long short-term memory (LSTM) models and a new hybrid model (CNN-LSTM), were employed to predict soil moisture at depths of 5 cm, 10 cm, 20 cm and 40 cm. Additionally, the impact of standalone sensor, texture features and multi-sensor data fusion on the accuracy of soil moisture prediction was explored. The results indicated that the model with RGB + Mul + TIR achieved the highest prediction accuracy, followed by those with Mul + TIR and RGB + Mul, with the coefficient of determination (R2) ranging 0.80–0.88, 0.64–0.84, and 0.60–0.81, and root mean square error (RMSE) ranging 2.46–2.99 m3·m−3, 2.86–3.89 m3·m−3 and 3.15–4.25 m3·m−3, respectively. Among single sensor inputs, the Mul sensor data has the highest prediction accuracy, followed by TIR and RGB sensor, with the coefficient of determination (R2) ranging 0.54–0.72, 0.36–0.52 and 0.14–0.26, and root mean square error (RMSE) ranging 3.72–4.58 %, 3.81–5.04 % and 4.27–6.21 %, respectively. The hybrid CNN-LSTM model exhibited the highest prediction accuracy, followed by CNN and LSTM models, with the coefficient of determination (R2) ranging 0.20–0.88, 0.16–0.83, and 0.14–0.81, and root mean square error (RMSE) ranging 2.46–5.01 m3·m−3, 2.68–5.35 m3·m−3 and 2.81–6.21 m3·m−3, respectively. The prediction accuracy of the models was the highest at the depth of 5 cm, followed by 10 cm, 20 cm and 40 cm, with the coefficient of determination (R2) average of 0.63, 0.62, 0.59, and 0.55, and root mean square error (RMSE) average of 3.70 m3·m−3, 3.79 m3·m−3, 3.85 m3·m−3 and 4.21 m3·m−3, respectively. Therefore, the hybrid CNN-LSTM model with RGB + Mul + TIR is recommended to predict soil moisture in citrus orchard. It provides method and data support for regional precision irrigation decision-making. • Multi-modal UAV sensor data was used to predict soil moisture at different depths. • The gradient vanishing and sequence fragmentation in transitional RNN was addressed. • The hybrid CNN-LSTM model outperformed CNN and LSTM models. • The prediction accuracy of the model was highest under RGB + Mul + TIR data fusion. [ABSTRACT FROM AUTHOR]

• A new hybrid PSO-LSTM model was proposed for soil moisture estimation. • The hybrid PSO-LSTM model had strong stability, universality and high prediction accuracy. • The prediction accuracy of the hybrid PSO-LSTM model was better than the LSTM model. • The prediction accuracy of models with vertical polarization was higher than that with cross-polarization. Soil moisture content is a vital variable in agricultural, hydrological, ecological and climatological processes. However, susceptible to soil type, soil structure, topography, vegetation and human activities, soil moisture content exhibits strong spatial heterogeneity in spatial distribution, which makes it difficult to accurately estimate the soil moisture content distribution information at a large scale using conventional methods. To solve the problem, this study proposed a novel hybrid model (PSO-LSTM) based on the particle swarm optimization (PSO) and long short-term memory (LSTM) network model to accurately predict soil moisture content at a large scale. Five different input combinations were constructed based on the vertical polarization (VV) and cross-polarization (VH) of multi-phase Sentinel-1A data, and the soil moisture content at depths of 5 cm, 10 cm, 20 cm and 40 cm in citrus orchards were estimated using the standalone LSTM and hybrid PSO-LSTM models. The results showed that the estimation accuracy of the hybrid PSO-LSTM model was greater than that of the standalone LSTM model at different depths, with the normalized root mean square error (NRMSE) of 4.568–11.023 % and 18.056–30.156 %, respectively. With the VV polarization as the only inputs, the PSO-LSTM model obtained high prediction accuracy, with the normalized root mean square error (NRMSE) of 5.458–10.125 %, respectively. Therefore, the PSO-LSTM model with VV polarization input was recommended to estimate the soil moisture content at different depths in citrus orchards, which provides important data for decision-making on distributed precision irrigation at a large scale. [ABSTRACT FROM AUTHOR]

Soil moisture is a significant variable in agricultural study and precision irrigation decision-making. It determines the soil water availability for plants, directly influencing plant growth, yield and quality. Owing to the variations in regional microclimate, landform difference, soil type and vegetation coverage, the soil moisture has strong spatial-temporal heterogeneity on a large regional scale. Micro-wave remote sensing can be used to invert soil moisture based on the dielectric constant under different weather conditions, while optical remote sensing utilizes spectral characteristics to estimate the physiological and ecological information of vegetation. In this study, two new hybrid models (ACO-RF and SSA-RF) were structured by optimizing the standalone random forest (RF) based on the ant colony optimization algorithm (ACO) and sparrow search algorithm (SSA), and six input combinations based on the multi-temporal Sentinel-1 and Landsat-8 remote sensing data from different sensors (optical, thermal and radar sensors) were used. The standalone RF, ACO-RF, and SSA-RF models with different combinations of inputs were employed to predict the soil moisture at different depths (5 cm, 10 cm, 20 cm, 40 cm) in a large-scale drip-irrigated citrus orchard. The results showed that the ACO-RF and SSA-RF outperformed the standalone RF model in terms of prediction accuracy at a depth of 0–40 cm, with R2 of 0.800–0.921 and 0.504–0.798, RRMSE of 7.214–16.284% and 11.124–22.214%, respectively. In the hybrid model, the ACO-RF model had better prediction accuracy than the SSA-RF model, with R2 of 0.805–0.921 and 0.800–0.911, RRMSE of 7.214–13.244% and 8.274–16.284%, respectively. At depths of 5 cm, 10 cm and 20 cm, the inversion accuracy of the model with microwave inputs was higher than that with multispectral inputs, with R2 of 0.556–0.888 and 0.541–0.886, RRMSE of 9.015–19.544% and 9.124–22.214%, respectively. However, at a depth of 40 cm, the inversion accuracy of the model with multispectral inputs was higher than that with microwave inputs, with R2 of 0.532–0.841 and 0.508–0.831, RRMSE of 9.124–21.021% and 9.142–21.214%, respectively. The model with multispectral, thermal, and microwave inputs exhibited the highest accuracy in predicting soil moisture, with R2 of 0.635–0.921, RRMSE of 7.214−18.564%, respectively. Therefore, the ACO-RF with multisource remote sensing data is recommended to predict the soil moisture in the drip-irrigated citrus orchard. This approach can provide data support for making intelligent irrigation decisions on a large-scale grid land lots. • Two new hybrid RF models were proposed for soil moisture prediction in a citrus orchard. • The hybrid model had better prediction accuracy than the standalone RF model at depths of 0–40 cm. • The model with multi-spectral, thermal and microwave inputs achieved highest prediction accuracy. • Models with microwave outperformed models with multi-spectral inputs at 5 cm, 10 cm and 20 cm depths, while the inversion accuracy with multi-spectral was better than that with microwave inputs at 40 cm depth. [ABSTRACT FROM AUTHOR]

• The canopy resistance was significantly correlated with R n , while it was correlated with LAI, θ and VPD. • The accuracy of estimated maize canopy resistance was improved by considering LAI threshold segmentation method. • The performance of Differential Evolution Algorithm was better than Genetic Algorithm. Accurate estimation of maize evapotranspiration (ET) is of great significance for the improvement of crop water use efficiency and precision irrigation. The Penman-Monteith model (P-M) has been widely used to simulate crop ET. In the P-M model, the estimation accuracy of canopy resistance (r c) has a direct impact on ET. In this study, based on the eddy covariance system, large-scale lysimeter and meteorological station data from three sites (Yucheng, Yangling and Shangqiu) in semi-humid regions of northern China, the P-M model was applied to obtain canopy resistance (r c-PM) and correlation significances between r c-PM and different impact factors (R n net radiation, T temperature, VPD saturated vapor pressure difference, θ soil moisture content, LAI leaf area index) were analysed. The whole growth period of maize was divided according to different LAI thresholds (0.1, 0.5, 1.0, 1.5, 2.0 and 3.0 m2 m−2). The Genetic Algorithms (GA) and Differential Evolution (DE) algorithms were used to optimize the empirical parameters of the Jarvis model, and the P-M model was applied to estimate ET under different LAI thresholds at the three stations. The correlation significances of r c-PM with different influencing factors followed the order R n > LAI > θ >VPD >T, and it was extremely significant with R n (P < 0.01) and significant with LAI and θ (P < 0.05). The GA and DE algorithm optimization results showed that the calculation accuracy of r c was highest when LAI=0.5 m2 m−2 at Yucheng station, with R2 of 0.80 and 0.81, respectively, and when LAI=1.0 m2 m−2, and the accuracy of r c was highest at Yangling station, with R2 of 0.87 and 0.89, respectively, and when LAI = 1.0 m2 m−2, and the accuracy of r c was highest at Shangqiu station, with R2 of 0.84 and 0.84, respectively. Combined with P-M model to calculate maize ET under different LAI thresholds, the simulation accuracy of ET was best when LAI = 0.5 m2 m−2 at Yucheng station, with averege R2 of 0.85, the order of simulation ET accuracy was: 0.5 >1.0 >1.5 >2.0 >3.0 >0.1 m2 m−2. When LAI = 1.0 m2 m−2, and the accuracy of maize ET was highest at Yangling and Shangqiu stations, with averege R2 of 0.83 and 0.85, respectively, the order of simulation ET accuracy was: 1.0>0.5>1.5>2.0>3.0>0.1 m2 m−2. ET accuracy calculated by the DE optimization algorithm was better than that of the GA optimization algorithm, with R2 of 0.40–0.84 and 0.58–0.86, respectively. This study suggests that the algorithm is of great importance to optimize the empirical parameters of the Jarvis model, of which DE optimization algorithm is recommended to simulate maize ET in semi-humid regions of northern China. [ABSTRACT FROM AUTHOR]

[Display omitted] • Extracting pure kiwifruit canopy information improves SMC estimation effectively. • The optimal band combination algorithm efficiently screens band and FVC variables. • The ensemble learning model for estimating SMC at different root-zone depths is developed. • Planted-by-planted mapping of SMC has potential for field irrigation management. Accurate and real-time monitoring of soil moisture content (SMC) is of utmost importance for effective field irrigation and maximizing crop water productivity. However, a comprehensive investigation into the inversion study for determining suitable combinations of unmanned aerial vehicle (UAV) image features and enhancing the precision of SMC model prediction has yet to be fully validated within a kiwifruit orchard setting. This study addresses this gap by employing a pre-processing method and an optimal band combination algorithm to assess the impact of various combinations of kiwifruit canopy reflectance and fraction vegetation coverage (FVC) features on the sensitivity of root-zone SMC. Furthermore, an optimal ensemble learning (EL) framework was developed to monitor SMC at various root-zone depths (0–10 cm [SMC10], 0–20 cm [SMC20], 0–30 cm [SMC30], 0–40 cm [SMC40], 0–50 cm [SMC50], 0–60 cm [SMC60]). The key findings of this research highlight the successful derivation of 10 wavebands and FVC features, exhibiting a strong correlation with SMC at different root depths. The gradient boosting (GBDT) model demonstrated the exceptional accuracy in estimating SMC10, with an impressive R2 value of 0.963 ± 0.030 and low RMSE values of 0.238 ± 0.111. Similarly, the eXtreme Gradient Boosting (XGBoost) model outperformed in estimating SMC20 to SMC60, with R2 and RMSE values of 0.963 ± 0.024 and 0.117 ± 0.053, respectively. Additionally, the utilization of the optimal EL model allows for digital mapping of SMC at different depths across fruit growth stages, showcasing superior adaptability for SMC30 to SMC60 (with R2 and RMSE of 0.782 ± 0.090 and 0.037 ± 0.011) compared to SMC10 and SMC20 (with R2 and RMSE of 0.765 ± 0.097 and 0.056 ± 0.024). These results underscore the potential of the EL estimation framework in characterizing the spatial distribution of root-zone SMC at the individual kiwifruit plant level. [ABSTRACT FROM AUTHOR]

Accurately estimating of soil moisture content (SMC) is essential for effective irrigation water management and optimizing plant water productivity. Recent advancements in multi-sensor platforms and ensemble learning (EL) algorithms such as the use of UAV-based multispectral (MS) and thermal infrared (TIR) images, have enabled precise mapping of SMC at the field level. The purpose of this study is to optimize the multi-dimensional (n-D) index set and EL model to achieve accurate inversion mapping of SMC in the 0–60 cm root zone at a kiwifruit orchard during the fruit growth period (May-Sep 2022). The n-D index set was computed using kiwifruit canopy MS and TIR data from UAV along with in-field temperature data (air temperature [Ta], kiwifruit canopy temperature [Tc], soil temperature [Ts], and Ta-Tc, Ta-Ts, Tc-Ts). The primary findings of this study are as follows: (1) The Three-Dimensional Drought Index (TDDI), Temperature Vegetation Drought Index (TVDI), and Crop Water Stress Index (CWSI) demonstrated superior performance in capturing temporal variations of root zone SMC compared to Normalized Difference Vegetation Index (NDVI) and Enhance Vegetation Index (EVI). (2) The optimal Categorical Boosting (Catboost) model with two optimal TDDIs (TDDI22: (Tc-Ts)-EVI10,6,2-CWSI space and TDDI7: Tc-NDVI7,3-CWSI space), exhibited exceptional performance for SMC estimation, with determination coefficient (R2) and root-mean-square error (RMSE) of 0.966 ± 0.010 and 0.046 ± 0.003%, respectively. (3) The planted-by-planted mapping performed well in estimating root zone SMC at different growth stages and under different irrigation treatments, with a correlation coefficient (R) of 0.936 ± 0.074 (p < 0.001). This study demonstrated that the integration of multi-sensor indicators and EL models can improve the accuracy of SMC estimation due to its robust adaptability to various field conditions. Furthermore, the planted-grid-based SMC mapping framework can be considered as a valuable approach for monitoring dynamic SMC, facilitating informed irrigation decisions at the individual planting grid. [Display omitted] • The n-D index set of SMC is successfully developed through multi-sensor data. • The TDDI set is more adaptive to temporal and spatial variations of SMC estimation. • The Boosting model demonstrates remarkable performance in estimating SMC. • The planted-by-planted mapping of SMC has potential for field irrigation application. [ABSTRACT FROM AUTHOR]

Accurate estimates of plant transpiration (T sf) are essential to maximizing the efficient use of water. When only considering the surface canopy resistance (r c), the Penman-Monteith (PM) model is commonly used in T sf modeling. However, the r c is difficult to measure but it can be accurately estimated using the Jarvis canopy resistance model (JA). Our objectives were to evaluate the r c of apple trees calculated with an inverted PM model integrated with different versions of the JA model for different growth stages and to compare their accuracy using four optimization algorithms, the least squares method (LSM), genetic algorithm (GA), quantum particle swarm optimization (QPSO), and the mind evolutionary algorithm (MEA). We explored the effect of environmental constraint functions and parameter optimization of three environmental variables, vapor pressure deficit (VPD), net irradiance (R n), and air temperature (T a), and of soil water content (θ a) on the accuracy of JA sub-models to calculate apple tree r c and T sf. In our analysis, we used rainfed data from experiments on an apple orchard conducted during 2008–2010 at Wuwei City on the Loess Plateau of China. We compared 81 segmented JA sub-models (by canopy growth stage), comprising of the combination of the environmental constraint functions that were used to calculate r c. Moreover, the JA sub-models were optimized and results were compared to improve the accuracy of T sf estimates with five empirical r c models, combined with the PM model. The results showed that sub-model JA 3322 , i.e., the third constraint for VPD, the third constraint for R n , the second constraint for T a , and the second constraint for θ a attained the best estimate of r c with a coefficient of determination (R2 = 0.71), a Nash-Sutcliffe efficiency coefficient (NSE = 0.65), root mean squared error (RMSE = 1257.4 s m-1), and global performance indicator (GPI = 1.0) for the whole growth stage. The equivalent values for the partial canopy stage were 0.78, 0.72, 1203.3 s m-1 and 0.99, and the values for the dense canopy stage were 0.78, 0.77, 445.6 s m-1 and 0.97, respectively. Segmented JA models based on the leaf area index threshold significantly improved the accuracy of r c estimation, where the median R2 and NSE were improved by 7.1 % and 12.4 % in the partial canopy stage and by 12.2 % and 13.4 % in the dense canopy stage. Despite pointing out the best environmental constraint functions of the JA model in the different growth stages, results indicated that the MEA yielded the most accurate estimates of r c , followed by QPSO, GA, and LSM. Moreover, the JA model with environmental constraints was the most accurate method to estimate the apple tree T sf , and MEA was the most suitable parameter optimization algorithm. Overall, the findings of this study provide accurate actual water consumption information of apple trees using easily accessible meteorological data for the effective day-to-day water management decision-making of rain-fed apple tree orchards on the Loess Plateau of China previously. • Jarvis was superior to other empirical models for canopy resistance estimation. • Jarvis segmented by canopy growth stage gave a more accurate estimation. • Finding the most appropriate Jarvis model for apple tree canopy resistance. • Intelligence optimization algorithms improved the accuracy of the Jarvis models. [ABSTRACT FROM AUTHOR]

[Display omitted] • Three types of r c models were embedded to single-source and dual-source ET models. • The r c coupled models optimized by GA performed better than original ET models. • The r c and ET were most sensitive to soil moisture and net radiation. • The SW-JA model performed best in modeling maize ET over the whole growing season. The reliability and accuracy of canopy resistance (r c) are crucial for the evapotranspiration (ET) estimation in sparsely vegetated ecosystems, but most studies are still limited to the single-source Penman-Monteith (PM) model. In this study, we coupled three types of r c models (three-way model: Jarvis, JA; two-way model: Kelliher-Leuning, KL; Farias, FA; one-way model: Katerji-Perrier, KP; Massman, MA) to the PM and Shuttleworth-Wallace (SW) models to assess their applicability and stability under single-source and dual-source hypotheses in a maize field in northwest China. The models were calibrated and validated against the observed data from a three-year (2011–2013) field experiment based on genetic algorithms (GA). The results indicated that the r c coupled models optimized by GA greatly alleviated the overestimation of the original PM and SW models with mean absolute error (MAE) decreased by 13.85%-40.77% and determination coefficient (R2) increased by 0.40%-7.37%. The SW-type models achieved overall high performance to simulate ET at a diurnal timescale over the entire crop season compared to the equivalent PM-type models. The three-way JA model including crop, soil and meteorological factors tended to perform more accurately than the other r c models, and the coupled SW-JA model was optimal for predicting maize ET over the entire growing season. The two-way KL model produced lower errors than the FA model, whereas the FA model had the potential to be improved due to the highest R2 with the field-based ET for both PM- and SW- hypotheses. In addition, the sensitivity analysis to environmental variables showed the effect of leaf area index (LAI) and vapor pressure deficit (VPD) was less on ET compared with net radiation (R n) and soil moisture (θ). This study contributes to the knowledge in the area of actual ET modeling over sparse vegetation, which provides a reference for agricultural and environmental applications. [ABSTRACT FROM AUTHOR]

Quantitatively characterizing and accurately predicting plant transpiration are of great significance, but directly measuring transpiration is impractical, time-consuming, and labor-intensive. This study compared the transpiration estimation performance of multiple linear regression (MLR), modified Jarvis–Stewart (MJS), and Shuttleworth–Wallace (S-W) with deep belief network (DBN), long short-term memory recurrent neural network (LSTM-RNN), and LSTM-RNN improved with multiple restricted Boltzmann machines (R- L -RNN) using 31 input combinations comprising complete subsets of Vapor pressure deficit (VPD), Net solar radiation (R n), Average air temperature (T a), Soil water content (SWC), and Leaf area index (LAI) observations collected at Wuwei, Changwu and Taigu stations on the Loess Plateau in China. The results showed that R- L -RNN obtained the most accurate estimations in the partial canopy stage, dense canopy stage, and whole growth stage, compared to MLR, MJS, S-W, DBN, and LSTM-RNN. The accuracy of the deep learning models (DNN) increased exponentially as the number of input variables increased, and the importance of the input variables followed the orders of: LAI > VPD > R n > T a > SWC in the partial and whole canopy stage, and VPD > R n > T a > LAI > SWC in the dense canopy stage. The apple tree transpiration models were more accurate in the partial and dense canopy stages than the whole growth stage. The coefficient of determination and Nash-Sutcliffe efficiency coefficient for the R- L -RNN model increased by 8.1–13.1% and 11.2–25.4% in the partial canopy stage, respectively, and by 2.6–6.9% and 14.7–20.1% in the dense canopy stage, whereas the relative root mean square error decreased by 8.7–28.6% and 17.3–38.2%. Overall, R- L -RNN is the most recommended model for estimating the apple tree transpiration, because it is such a simple method that agricultural water managers can easily determine the water consumption of apple trees using limited accessible observational data. • A hybrid deep learning model provided most accurate apple tree transpiration model. • Importance of different inputs to deep learning models differed with growth stages. • Accuracy of deep learning models improved exponentially with more inputs. • The transpiration models were more accurate in the partial and dense canopy stages. [ABSTRACT FROM AUTHOR]