A Comparative Study of Pretrained Vision Transformer and VGG16 Models for Soil Moisture Classification

Accurate soil moisture monitoring is of prime importance to precision agriculture, water resource management, as well as for environmental monitoring. Traditional methods often suffer due to limited spatial coverage and manual data collection. Recent advancements in computer vision have significantly enhanced the ability to comprehend and analyse image data. These technological breakthroughs have made it possible for accurate and efficient image classification to find its way into computer vision as well as the realm of machine learning. This study introduces a comparative analysis of soil moisture classification using a pre-trained Vision Transformer (ViT) and VGG16 models using high-resolution RGB images to classify dry and wet soil. The model was trained on a new dataset of 1028 images of soils taken from agricultural fields and outdoor locations. VGG16 had excellent train, validation, and test accuracies at 98.42%, 96.41%, and 95.96%, respectively, and was superior to the ViT model at 97.86%, 90.41%, and 91.13%. This is due to the ability of the VGG16 model to more effectively capture local spatial features and also its robustness against input variations and noise within the soil image dataset. Moreover, the efficient inference time of 70 ms attributed to the VGG16 architecture makes it generally well-suited for real-time applications in precision agriculture and environmental sensing.