ViT-DARTS: A Parameter-Efficient, Vision Transformer-Based Differentiable Architecture Search for UAV-Based Sward Height Regression

Sward height is a crucial indicator in pasture management. Unmanned Aerial Vehicle (UAV) imagery is commonly used to estimate sward height through Digital Surface Models (DSMs) by subtracting ground and canopy elevation. However, this method requires substantial expertise to manually correct the inaccuracies caused by variable terrain and dense vegetation canopies. Deep learning methods for estimating sward height remain underexplored, primarily due to the labour-intensive collection of diverse field measurements and the computational demands required to train customized solutions for specific farm conditions from existing models. Additionally, manually designing model architectures is time-consuming and resource-intensive.This paper aims to automatically develop a parameter-efficient deep learning model for sward height estimation and other agricultural plant height regression tasks using UAV-derived DSMs. To achieve this, the ViT-DARTS method is introduced, integrating Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs) as candidates within the Differentiable Architecture Search (DARTS) framework, and layer group lasso regularization to promote sparsity and reduce overfitting.After training on a sward height dataset, we evaluated the performance of the ViT-DARTS-searched model against the original DARTS-searched model and existing CNN and ViT-based deep learning models across four diverse sward and crop datasets, using varying amounts of training data. The results demonstrated that the ViT-DARTS-searched model exhibited robust performance when trained from scratch under different datasets, and using 3,900 times fewer parameters than the classical CNN-based model, AlexNet.