Detection of Moss Coverage and Growth Rate on Rock Surfaces Based on Semantic Segmentation

Mosses are one of the most widespread and dominant biological diseases that damage stone cultural relics. This has resulted in discoloration, cracking, and even collapse of rocks, which have severely impacted the inheritance of outstanding traditional cultures. The basis of effective preventive conservation is to determine the nondestructive and accurate detection of key parameters related to the development of biological diseases on stone cultural relics. Therefore, to detect the moss coverage rate and growth rate, we constructed a semantic segmentation model for moss segmentation on rock surfaces. Subsequently, the local and texture information extracted by the shallow structure of the convolutional neural network was combined with Swin Transformer to enhance feature diversity. Next, transfer learning was performed on similar large-scale datasets. The effectiveness of the model was validated using a public dataset and the moss dataset. Our proposed model can accurately identify the moss coverage rate and growth rate on rock surfaces, achieving an average intersection over union of 53.05 and 94.79% on the LoveDA dataset and moss dataset, respectively. The research results provide important support for the preventive conservation of stone cultural relics.