Crop Yield Prediction Using Remote Sensing Space Convolutional Attention Machine Learning with Potter Optimisation Algorithm for Independent Variable Selection

Abstract

Accurate crop yield prediction is crucial for food security planning, resource allocation, and agricultural policy formulation. This paper proposes a novel hybrid deep learning framework, the Space Convolutional Attention Machine Learning (SCAML) model, that integrates multi-source remote sensing imagery with a biologically-inspired Potter Optimisation Algorithm (POA) for principled independent variable selection. The proposed model combines a dual-branch spatial-spectral Convolutional Neural Network (CNN) for local feature extraction with a multi-head self-attention transformer for temporal dependency modelling, culminating in a Space Convolutional Attention Module (SCAM) that fuses both representations. The Potter Optimisation Algorithm, inspired by the nest-building behaviour of potter wasps (Eumenes spp.), performs wrapper-based feature selection to identify the most informative spectral indices and ancillary variables, significantly reducing dimensionality while preserving predictive power. Experiments conducted on multi-season rice, wheat, and maize datasets across five agroclimatic zones demonstrate that SCAML-POA achieves a Root Mean Square Error (RMSE) of 0.312 t/ha, a Mean Absolute Percentage Error (MAPE) of 3.47%, and an R² of 0.964, outperforming baseline CNN, LSTM, and Random Forest regressors. The results affirm that the synergy between optimised variable selection and spatiotemporal attention learning can substantially improve crop yield forecasting accuracy.