Machine Learning Models Based on UAV Oblique Images Improved Above‐Ground Biomass Estimation Accuracy Across Diverse Grasslands on the Qinghai–Tibetan Plateau

Abstract

ABSTRACTUnmanned aerial vehicles (UAVs) are becoming important tools for modern management and scientific research of grassland resources, especially in the dynamic monitoring of above‐ground biomass (AGB). However, current studies rely mostly on vertical images to construct models, with little consideration given to oblique images. Determination of image acquisition height often relies on experience and intuition, but there is limited comparison of models in estimating across different grassland types. To address this gap, this study selected 56 plots on the northern Qinghai–Tibetan Plateau (QTP), comprising 16 alpine meadows (AM), 14 alpine steppes (AS), 13 alpine meadow steppes (AMS), and 13 alpine desert steppes (ADS). We used the DJI Mavic 2 Pro to capture a total of 5040 images at six heights (5, 10, 20, 30, 40, and 50 m) and five angles (30°, 45°, 60°, 90°, and 180° panoramic shots). Based on RGB (red‐green‐blue) images, seven vegetation indices (normalized difference index (NDI), excess red vegetation index (EXR), modified green red vegetation index (MGRVI), visible atmospherically resistant index (VARI), excess green minus excess (EXG), green leaf index (GLI), and red–green–blue vegetation index (RGBVI)) were employed, displaying a trend in vegetation and biomass changes across different heights and angles, peaking at 20 m and 45°. Linear regression models and machine learning models (random forest, extreme gradient boosting, multilayer perceptron neural network, and stochastic gradient descent) were generated, with NDI, VARI, and MGRVI providing the best estimations. Comparative results on estimations of different grassland types indicated that oblique images helped reduce the models' root mean square error (RMSE), particularly in the machine learning models. All models were best in AMS and ADS, with average R2 of 0.810 and 0.825, with machine learning models (average R2 = 0.746) stronger than linear regression models (average R2 = 0.597), indicating specific requirements for model selection across different grasslands. The findings in this study can provide a reference for the adaptive management of different grassland ecosystems on the QTP and worldwide.