Abstract
This research introduces an innovative agricultural carbon accounting approach for straw burning that combines stochastic process modeling with LSTM neural networks. Traditional methods face limitations including high uncertainty, fragmented data, and prohibitive real-time monitoring costs. Our off-site inverse carbon accounting methodology employs three-dimensional Brownian motion to simulate carbon molecular diffusion patterns, incorporating horizontally drifted motion influenced by wind speed and vertically truncated motion dominated by thermal activity. The framework utilizes LSTM-based time-series predictions to generate virtual diffusion path samples for dynamic model calibration. By quantifying the probability density function of carbon molecular diffusion, we inversely derive carbon emission rates from particle arrival probabilities at observation points. Validation through a straw-burning case demonstrates an average carbon emission rate of 0.0049 tons/second with error margins below 10%, confirming the method's accuracy. This approach overcomes limitations of traditional emission factor methods while providing cost-effective real-time carbon monitoring for agricultural contexts. Future research could integrate multi-physics models, remote sensing data, and advanced computational techniques like quantum computing to enhance scalability and precision. This work establishes a foundation for data-driven carbon governance in agricultural supply chains, supporting global carbon neutrality efforts.