Аннотация
Accurate crop yield forecasting plays an important role in precision agriculture and sustainable food production. This study presents a hybrid Dense-LSTM neural network model for forecasting spring wheat yield under varying phosphorus application levels and meteorological conditions in Kazakhstan. The dataset was constructed using regional agricultural and meteorological observations collected between 2000 and 2022. To improve data quality and model stability, several preprocessing techniques were applied, including median-based missing value imputation, outlier handling and feature normalization using MinMax and StandardScaler approaches. The proposed hybrid architecture combined Dense and LSTM layers to capture nonlinear relationships and temporal dependencies within agricultural data. Model performance was evaluated using MSE, RMSE, MAE and R² metrics and compared with traditional machine learning approaches, including Linear Regression, Random Forest and Gradient Boosting models. Experimental results demonstrated that incorporating phosphorus-related indicators into the input feature space improved forecasting accuracy. Among the evaluated approaches, StandardScaler normalization combined with the proposed hybrid neural network achieved the best predictive performance with an MSE value of 7.36. The obtained results demonstrate the applicability of hybrid deep learning architectures for agricultural forecasting tasks under Kazakhstan-specific agro-climatic conditions and highlight the importance of integrating environmental and agronomic factors in precision agriculture applications.