NDVI Prediction with RGB UAV Imagery Utilizing Advanced Machine Learning Regression Models

Abstract. The ever-evolving technology has significantly affected the sensors used in UAV cameras and has played an important role in the expansion of the application areas of hobbyist and commercial UAVs. In particular, UAVs with multispectral (MS) cameras, which have the potential to detect a wide range of spectral information, are widely used in many popular research areas such as precision agriculture and forestry. However, despite their advanced capabilities, the high cost of these technologies limits their accessibility for basic users. In this study, the agricultural potential of RGB UAVs, which have a much wider user base due to their lower cost, was investigated by predicting the Normalized Difference Vegetation Index (NDVI), which is widely preferred for plant classification, growth and health monitoring. In the literature, RGB camera-based NDVI prediction studies involving machine learning and deep learning algorithms have focused on the correlation of the results with the reference data (R²) or the model accuracy of the algorithms used. The approaches applied have generally been tested on single photographs or solely on vegetation areas. In this study, using the MS UAV NDVI map as reference, a comprehensive evaluation approach was applied where each pixel of the NDVI prediction maps produced by categorical boosting (CatBoost), light gradient boosting machine (LightGBM) and a stacking ensemble learning model obtained from the combination of both algorithms, whose performance in NDVI estimation has not been tested extensively before. The models were tested in an urban area with numerous buildings and a large study area with dense vegetation. The performance of the NDVI maps was analyzed using R², Root Mean Square Error (RMSE), Normalized Median Absolute Deviation (NMAD) and Standard Deviation (STD) metrics. As a result of the comprehensive analysis, it was found that the models performed similarly in general, but the LightGBM model was slightly behind the others. The considerable results around 0.81–0.83 as R² and ~0.09 as RMS and STD clearly showed that RGB cameras can be a lower-cost alternative solution for generating NDVI maps in agricultural studies when supported by machine learning models.