Minimizing active canopy sensor differences in nitrogen status diagnosis and in-season nitrogen recommendation for maize with multi-source data fusion and machine learning

Active canopy sensors (ACSs) are great tools for diagnosing crop nitrogen (N) status and grain yield prediction to support precision N management strategies. Different commercial ACSs are available and their performances in crop N status diagnosis and recommendation may vary. The objective of this study was to determine the potential to minimize the differences of two commonly used ACSs (GreenSeeker and Crop Circle ACS-430) in maize (Zea mays L.) N status diagnosis and recommendation with multi-source data fusion and machine learning. The regression model was based on simple regression or machine learning regression including ancillary information of soil properties, weather conditions, and crop management information. Results of simple regression models indicated that Crop Circle ACS-430 with red-edge based vegetation indices performed better than GreenSeeker in estimating N nutrition index (NNI) (R-2 = 0.63 vs. 0.50-0.51) and predicting grain yield (R-2 = 0.56-0.57 vs. 0.49). The random forest regression (RFR) models using vegetation indices and ancillary data greatly improved the prediction of NNI (R-2 = 0.81-0.82) and grain yield (R-2 = 0.87-0.89), regardless of the sensor type or the vegetation index used. Using RFR models, moderate degree of accuracy in N status diagnosis was achieved based on either GreenSeeker or Crop Circle ACS-430. In comparison, using simple regression models based on spectral data only, the accuracy was significantly lower. When these two ACSs were used independently, they performed similarly in N fertilizer recommendation (R-2 = 0.57-0.60). Hybrid RFR models were established using vegetation indices from both ACSs and ancillary data, which could be used to diagnose maize N status (moderate accuracy) and make side-dress N recommendations (R-2 = 0.62-0.67) using any of the two ACSs. It is concluded that the use of multi-source data fusion with machine learning model could improve the accuracy of ACS-based N status diagnosis and recommendation and minimize the performance differences of different active sensors. The results of this research indicated the potential to develop machine learning models using multi-sensor and multi-source data fusion for more universal applications.