Abstract The goal of this study was to update the CSM‐CROPGRO‐Cotton model for its ability to simulate cotton ( Gossypium hirsutum L.) lint yield and cardinal temperature effects on physiological processes, followed by a thorough recalibration against extensive data on seed‐cotton yield, lint yield and percentage, crop mass, and time series growth data collected on 11 cotton cultivars in 143 treatments over 21 site‐years in Arizona, Texas, Georgia, and North Carolina (where CSM is Cropping System Model). Model parameters were updated for seed‐cotton composition, cardinal temperatures for critical processes, and dependency of lint percentage on temperature, drought and N stress, and cultivar. The D ‐statistic for time series variables averaged over all site treatments was 0.947, 0.977, 0.896, 0.865, 0.937, 0.978, 0.955, and 0.974 for canopy height, total crop mass, leaf area index, leaf mass, stem mass, boll mass, boll harvest index, and vegetative stage, respectively. Additional data from 46 treatments in 10 site‐years in Mississippi were used for calibrating lint yield and percentage. For these treatments, relative root mean square error (RRMSE) was 16.5%, 20.0%, 20.7%, and 2.7% for crop mass, seed‐cotton yield, lint yield, and percentage, while respective D ‐statistics were 0.871, 0.822, 0.786, and 0.795. For end‐of‐season outputs across treatments in all 31 site‐years of experiments, RRMSE was 14.5%, 25.9%, 22.4%, and 6.0% for crop mass, seed‐cotton yield, lint yield, and percentage, while respective D ‐statistics were 0.950, 0.925, 0.955, and 0.869. The model was successfully evaluated against lint yield and percentage for the 1999–2019 Arkansas cotton variety trials. The improved model, its parameters, and appropriate experimental files will be included in a future release of DSSAT (Decision Support System for Agrotechnology Transfer).
Abstract Accurate soil moisture prediction is essential for precision irrigation and improving water‐use efficiency in peanut ( Arachis hypogaea L.) production. This study developed a bidirectional transformer (BiTransformer)‐based soil moisture forecasting framework integrated with a growth‐stage irrigation decision model for peanut cultivation in Luanzhou, North China. Hourly meteorological and soil moisture data collected during the 2023–2025 growing seasons were used for model development and validation. The proposed model combines BiTransformer encoders, cross‐attention interaction, temporal attention, and gated feature fusion to capture complex temporal dependencies and nonlinear dynamics of soil moisture. Model performance was compared with random forest, XGBoost, recurrent neural network (RNN), long short‐term memory (LSTM), BiLSTM, and conventional Transformer models using multiple evaluation metrics. Results showed that the proposed BiTransformer achieved the highest prediction accuracy, with an R 2 of 0.9707, root mean squared error (RMSE) of 0.0144, and mean absolute error (MAE) of 0.0096, outperforming all benchmark models. Ablation experiments further demonstrated that both the bidirectional architecture and attention mechanism contributed substantially to model performance. Based on predicted soil moisture and peanut water requirement standards, a quantitative irrigation scheduling strategy was established for different growth stages. The decision model successfully identified irrigation timing and recommended irrigation amounts according to crop water demand, particularly during the podding stage when water requirements were highest. The proposed framework integrates soil moisture prediction with irrigation decision‐making and provides an effective tool for precision irrigation management, water conservation, and sustainable peanut production.
Abstract Additions of urease and nitrification inhibitors to nitrogen (N) fertilizers can reduce N losses and improve crops’ N use efficiency (NUE). However, the benefits of inhibitors depend on interactions among crops, soils, and weather variables. Root Zone Water Quality (RZWQM2) models can simulate such complex interactions to predict N dynamics but need evaluation for systems with additions of inhibitors. For this study, RZWQM2 model was used to simulate the dual applications of nitrification and urease inhibitors (Inhibitor) on the spring wheat grain yield and denitrification loss of N in the form of nitrous oxide (N 2 O) in western Minnesota during 2014 and 2015 growing seasons. The RZWQM2 model was calibrated for the urea only application at 146 kg N ha −1 . Model reliably reproduced spring wheat yield across years and treatments, with simulated values closely matching observation ( R 2 = 0.64, Nash–Sutcliffe model efficiency [NSE] = 0.49, relative root means square error [RRMSE] = 0.14, mean absolute percent error [MAPE] = 11.5). Model overestimated yield by 29% and flux by 326% under urea only treatment during dry season (2014) but consistently captured the relative reduction in denitrification N loss from inhibitor additions. Long‐term scenario simulations (2000–2024) indicated that inhibitors clearly reduced N losses and that wetter springs (+10%) decreased the probability of higher N 2 O emissions, while drier springs (−10%) increased the flux which is consistent with other related studies. This study highlights the long‐term potential of inhibitors to mitigate fertilizer‐N losses. It also demonstrates the utility of the RZWQM2 model in evaluating inhibitor efficacy under spring wheat production in the Northern Great Plains.
Abstract Accurate estimation of seed yield in perennial grains like Thinopyrum intermedium (intermediate wheatgrass [IWG]) remains a challenge due to high interannual and spatial variability of this important trait. We evaluated the potential of before‐harvest unmanned aerial vehicle (UAV)‐derived multispectral vegetation indices to predict IWG seed yield under experimental conditions in northern Utah over three consecutive years (2022–2024). A global Random Forest model integrating all years was compared against year‐specific models using standard metrics ( R 2 , root mean square error [RMSE], mean absolute error, and Lin's concordance correlation coefficient). The global model outperformed individual‐year models ( R 2 = 0.813; RMSE = 24.3 kg ha). Chlorophyll‐sensitive indices such as green normalized difference vegetation index and chlorophyll index green were ranked as the most influential predictors, while classic indices including normalized difference vegetation index and the ratio vegetation index showed limited predictive power. These findings highlight the effectiveness of pre‐harvest UAV imagery for yield prediction in perennial grasses and support the use of multi‐year model training frameworks.
Abstract During the tobacco ( Nicotiana tabacum L.) leaf flue‐curing process, various types of nonconforming tobacco leaves are often produced due to the influence of multiple factors, significantly reducing the usability value of cured tobacco leaves. Based on different formation mechanisms, defective tobacco leaves can be broadly categorized into eight types, including steamed slice, black‐rotten leaf, waterlogged vein/slice, greenish‐yellow leaf, mold‐damaged leaf, over‐reddened cured leaf, gray‐spotted leaf, and toxicant‐damaged leaf. Traditional manual sensory classification methods are highly subjective and error‐prone, hindering early detection and intervention of curing defects. To improve the accuracy, convenience, and automation of defective tobacco leaves sorting, a mobile intelligent recognition application integrating machine vision with machine and deep learning technologies was developed in this study. The application adopted an innovative dual‐mode architecture: an offline mode utilizing the random forest algorithm for operation without network connectivity with the classification accuracy of 85.8%, and an online mode employing the visual self‐supervised pretraining (VIMER‐CAE) model through cloud collaboration with 90.4% recognition accuracy. In addition to precise identification, the application combined cause analysis with preventive recommendations, requiring no specialized equipment or manual intervention beyond a smartphone. Compared to traditional empirical methods, this application replaces subjective judgment with machine vision analysis and provides process optimization guidance to reduce tobacco leaf quality losses. Its portable, high‐precision, and intelligent design offers an innovative solution for curing process control, helping tobacco farmers improve production efficiency and economic benefits while promoting the standardization and intelligent development of tobacco cultivation.
