A Data-Centric Approach for Wind Plant Instance-Level Segmentation Using Semantic Segmentation and GIS

Wind energy is one of Brazil’s most promising sources, and the rapid growth wind plants has increased need for accurate efficient inspection methods. The current onsite visits, which are laborious costly, have become unsustainable due to sheer scale across country. This study proposes a novel data-centric approach integrating semantic segmentation GIS obtain instance-level predictions by using free orbital satellite images. Additionally, we introduce new annotation pattern, includes turbines their shadows, leading larger object size. elaboration data collection used panchromatic band China–Brazil Earth Resources Satellite (CBERS) 4A, with 2-m spatial resolution, comprising 21 CBERS 4A scenes more than 5000 annotated manually. database 5021 patches, each 128 × dimensions. deep learning model comparison involved evaluating six architectures three backbones, totaling 15 models. sliding windows allowed us classify large areas, considering different pass values balance between performance computational time. main results from this include: (1) LinkNet architecture Efficient-Net-B7 backbone was best model, achieving an intersection over union score 71%; (2) use smaller stride improves recognition process areas but increases power, (3) conversion raster polygon in platforms leads highly predictions. entire pipeline can be easily applied mapping Brazil expanded other regions worldwide. With approach, aim provide cost-effective solution inspecting monitoring plants, contributing sustainability sector beyond.