Integrating Machine Learning and Remote Sensing to Determine Crop Nitrogen Content of Maize

Highlights Crop nitrogen content (CNC) of maize was predicted using UAVs and satellite-based VIs and ML models. Predicted CNC was validated using Landsat 8 data against the actual field values. RF model with a combination of ENDVI, NDVI, and SAVI found to be effective in predicting CNC. Soil background significantly affects the spectral indices and influences the CNC. ABSTRACT. Predicting crop nitrogen content (CNC) of maize in real-time is crucial for effective nitrogen (N) management and the overall fertilizer application plan. The objective of this research was to identify a suitable machine learning (ML) model for assessing the CNC of maize. This approach used vegetation indices (VIs) derived from unmanned aerial vehicles (UAVs) and satellite (Landsat-8 and Sentinel-2) imagery to obtain real-time spatio-temporal information on CNC at field and regional scales, thereby developing in-season N management tools. Various spectral vegetative indices (VIs) were extracted from UAV images. These indices are used as input parameters for distinct ML models such as the Random Forest algorithm (RF), Support Vector Machine (SVM), Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), and a linear regression model. Statistical parameters such as mean absolute error (MAE), coefficient of determination (R 2 ), Root Mean Square Error (RMSE), Nash-Sutcliffe efficiency (NSE), percent bias (PBIAS), Kling–Gupta efficiency (KGE), and mean absolute percentage error (MAPE) were used in a randomized 10-fold cross-validation technique to evaluate the performance of the models. All these models were developed to predict CNC at the field scale, and, further, the applicability of the best model was tested at the regional scale across different growth stages. The findings of the study revealed that the RF algorithm exhibited superior performance, with RMSE, MAE, R 2 , NSE, PBIAS (%), MAPE (%), and KGE of 0.58, 0.41, 0.64, 0.63, 1.37, 17.24, and 0.67 for the testing of CNC prediction. Furthermore, it was shown that the performance of the model was more significantly influenced by the inclusion of VIs compared to individual spectral bands. This study observed that the RF model is effective in predicting CNC with the VIs combination of ENDVI, NDVI, and SAVI for Landsat-8 and Sentinel-2 images. The comparison of actual and predicted CNC using satellite imagery showed a similar trend for the different months, except during the initial growth stage. While these results offer valuable insights, our study indicated that the spatial resolution of satellite imagery can significantly influence the accuracy of CNC prediction. During the early growth stages of maize, limited canopy coverage and increased soil background exposure further reduce model reliability. Overall, this indicates the potential to assist farmers by precise prediction of spatio-temporal variation in CNC during different growth stages utilizing ML models.