HECYO-NET: HYBRID ENSEMBLE MOTH FLAME AND SHARK SMELL OPTIMIZATION MODEL FOR CROP MANAGEMENT

Abstract

Precision agriculture needs accurate, scalable and adaptive models to handle issues like noisy data, sensitivity to outliers, heterogeneity and generalizability of crop yield prediction. Conventional machine learning and deep learning (DL) algorithms tend to be unreliable because of environmental conditions and inconsistencies in large-scale data. To address these issues, this paper proposes the Hybrid Evolutionary Crop Yield Optimizer Network (HECYO-Net), which combines state-of-the-art preprocessing, feature engineering, and a new Hybrid Flame-Shark Evolutionary Optimizer (HFSEO). The framework uses K-Nearest Neighbours (KNN) based imputation with robust scaling as a preprocessing step, denoising autoencoders to reduce noise and extract latent features and ReLU-based ensemble construction trained with Root Mean Square Propagation (RMSProp). The optimisation is carried out by HFSEO, a hybrid of Moth Flame Optimization (MFO) and Shark Smell Optimization (SSO) to provide convergence and escape local minima. Deep Q-Networks (DQN) are also applied to make adaptive agricultural recommendations. Experimental results on a large dataset of 1,000,000 agricultural records demonstrate that HECYO-Net obtains the best performance, MAE = 0.09, MSE = 0.05, RMSE = 0.14, MAPE = 0.09, and nRMSE = 0.28, compared to CatBoost, XGBoost, LightGBM, CNN-RNN, and other ensemble baselines. The results demonstrate the importance of the model in delivering sustainable, precise and computationally effective crop yield prediction and decision support to the farmers.