Keep the key part: exploring drone-captured digital elevation model data augmentation for deep learning-based crop height estimation

In precision farming, accurate crop height estimation is crucial. Deep neural networks (DNNs) have been shown to effectively estimate crop height from drone-captured images, offering more precise and automated measurements. However, the reliability of RGB and multispectral images is compromised by lighting conditions, shadows, and seasonal variations. Digital Elevation Models (DEMs) derived from drone images provide a stable terrain elevation representation, which is less susceptible to environmental fluctuations. Moreover, training DNN models demands a large amount of data, necessitating extensive datasets for each crop height measurement setting. Data augmentation is a common strategy to enhance data diversity. It works well on RGB images by mimicking real-world variations to help the models adapt to a range of environments. However, applying similar augmentation methods to DEM can alter its geometry, potentially misrepresenting the terrain elevation and introducing errors and noise. This study investigates the performance of augmentation methods on DEM and introduces the KeepKey DEM augmentation, which is designed to preserve crucial DEM features while increasing data diversity, thus improving DNN training efficiency without additional data collection and keeping data variation within a controllable level. Our evaluation, conducted on a public cotton dataset from a sloped area, demonstrates that the KeepKey DEM augmentation improves DNN performance, reducing Root Mean Square Error by up to 19.36% compared to non-augmented models, by 39.1% compared to AutoAugment, by 14.29% compared to models with conventional augmentation method, by 23.5% compared to DEM-specific augmentation- Terrain amplification, by 40.7% compared to manual ground-canopy differentiation.