Machine Learning Predicted Crop Yield and Soil Organic Carbon Variation After Biochar Application

ABSTRACT The prediction of crop yield is of significant interest for growers utilising biochar on land, while managers prioritise the enhancement of soil organic carbon (SOC). However, it remains unclear to what extent machine learning can accurately predict crop yield or SOC when biochar is applied to soil. In this study, Random Forest (RF) and Multilayer Perceptron Neural Network (MLP‐NN) models were employed to predict crop yield and SOC with 297 paired data from field trials. The results indicated that the RF model (test R 2 = 0.83) did not differ significantly from the MLP‐NN model (test R 2 = 0.84) in predicting crop yield. However, the RF model (test R 2 = 0.87) performs significantly better than the MLP‐NN model (test R 2 = 0.53) in predicting SOC. The most influential features for crop yield were found to be the biochar application rate (15%), initial SOC (13%), biochar pH (10%), and biochar TP (10%). In contrast, the variation of SOC was primarily influenced by latitude (26%), biochar application rate (22%), initial SOC (15%), and biochar pH (13%). Furthermore, both crop yield and SOC variation were influenced by multiple factors, not solely one, and their impacts were not necessarily linear. This study suggests that the optimization of biochar pH and phosphorus content, along with the regulation of its application rate in sandy or clay‐rich soils, can simultaneously enhance both crop yield and SOC. In the future, we hope to develop a decision support system with prediction, different scenarios, and consultation capabilities based on geospatial location.