Integrating Data Envelopment Analysis and Machine Learning Approaches for Energy Optimization, Decreased Carbon Footprints, and Wheat Yield Prediction Across North-Western India

An accurate, timely and large-scale yield prediction is considered critically important to frame policies to mitigate climate change risks and for ensured food security. Wheat (Triticum aestivum L.), being a major cereal grain crop of north-western India, has been highly energy intensive with large emissions of greenhouse gases, affecting ecosystems’ sustainability. We therefore, performed energy optimization in wheat for enhanced ecosystems’ resilience and carbon (C) sustainability, while predicting yields of 120 intensive wheat production systems across north-western India. An integrated approach of data envelopment analysis (DEA) and machine learning (ML) models was applied for energy optimization and accurate yield prediction. After optimization of energy input in wheat production, 8 different ML models of variable complexity viz. linear regression (LR), ridge regression (RR), lasso regression (LAR), elastic net regression (ENR), decision tree regression (DT), random forest regression (RF), gradient boosting regression (GB) and support vector regression (SVR) were applied and evaluated statistically for accurate prediction of wheat yields of studied decision making units (DMUs). These results revealed that wheat ecosystems have a high total input energy (EI) of 24.7 GJ ha−1, which comprises ~46.5% share of direct and 53.5% of indirect energy sources. The non-renewable energy input was ~3.9 times higher than the renewable energy. The total C equivalent emissions of 1815.7 kg CO2e ha−1 comprised the highest share of chemical fertilizers (~49.8%), and exhibited a linear significant relationship with EI (R2 = 0.8117**; p < 0.01). Nitrogenous fertilizers application contributes ~90.1% towards total fertilizer related energy input in wheat production. The net global warming potential of 2.30 Mg CO2e ha−1 yr−1 was estimated which resulted in yield scaled emissions (i.e., greenhouse gases intensity) of 0.34 kg CO2e kg−1 wheat grain yield. The DEA-based benchmarking showed that the technical efficiency (TE) score of 79 DMUs was < 1.00, elucidated ~65% DMUs as energy inefficient. The average (n = 120) TE score of 0.92 illustrates energy saving possibility of 2464.6 MJ ha−1 (~8% of EI), mostly via efficient fertilizer (54.4%) and irrigation water management (11.6%). The 1:1 (x = y) correspondence analysis implies that DT, GB, and RF models can accurately predict wheat productivity with significantly higher R2 values of 0.998** (p < 0.01), 0.950** (p < 0.01), and 0.832* (p < 0.05), respectively. These results underpin overwhelming significance of DEA-based energy optimization for wheat ecosystems, which helps reduce the energy and C footprints for sustainable and cleaner production. Nonetheless, DT, GB, and RF models outperformed in yield prediction with the lowest root mean square error and normalized mean square error, but the highest Nash–Sutcliffe efficiency and index of agreement. Therefore, DEA-based benchmarking has significant energy saving potential, while DT, GB, and RF models have highly accurate wheat yield prediction abilities for north-western India.