Abstract—This research investigates the development and evaluation of self-supervised learning models for groundnut yield forecasting, aiming to support precision agriculture through accurate, data-driven predictions. The study focuses on three model architectures within a self-supervised framework: a Convolutional Neural Network (CNN) for spatial data, a Recurrent Neural Network-Transformer (RNN-Transformer) hybrid for temporal data, and a Graph Neural Network (GNN) for relational agronomic data. Each model is trained using a tailored pretext task masked patch reconstruction for CNN, time-series forecasting for RNN-Transformer, and link prediction for GNN enabling the extraction of meaningful features from unlabeled datasets. The models were trained on multisource data, including Sentinel-2 satellite imagery, European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5) dataset provided by the ECMWF and India Meteorological Department (IMD) weather data, and soil-agronomic records from ICRISAT and Soil Grids. Comparative results show that the GNN-based model achieved the best performance, with a Root Mean Square Error (RMSE) of 176.4 kg/ha, Mean Absolute Percentage Error (MAPE) of 6.3%, and R² of 0.93. In contrast, the CNN and RNN–Transformer models reported higher RMSE values (245.8 kg/ha and 218.3 kg/ha) and lower R² scores (0.82 and 0.87), confirming the superior predictive accuracy of the GNN approach. The GNN also demonstrated strong regional generalization, achieving an R² of 0.93 in the Southern Semi-Arid Zone, and showed superior pretext task accuracy at 93.6%. Additionally, it required only 58 min of training and converted into 22 epochs, offering a balanced profile of accuracy and efficiency. These findings confirm the effectiveness of graph-based, self-supervised learning in modeling complex agricultural systems and highlight its potential for scalable deployment in real-world precision agriculture applications.
 
Keywords—self-supervised learning, yield prediction, precision agriculture, Graph Neural Network (GNN), remote sensing, deep learning architectures


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).