Investigation on Landfill Segmentation Performance from Remote Sensing Images using Attention Mechanism-Improved YOLOv7 Model

Yating Zhou, Long Chen, Yongsheng Zhou, Fan Zhang

Abstract: Landfill monitoring is of paramount importance for environmental sustainability and public health. Remote sensing technology enables the acquisition of high-resolution spatial data over large areas, providing a comprehensive view of landfill sites. The integration of deep learning, particularly convolutional neural networks, enhances landfill segmentation accuracy by automating feature extraction and recognizing complex spatial patterns. However, challenges persist in obtaining labeled training data and identifying specific landfill details. To track these challenges, this paper proposes an advanced method for landfill segmentation. Firstly, we improve the YOLOv7 model by introducing the attention mechanism and pre-training it with the dataset of 313 remote-sensing images of landfills. Then, wavelet transform, and Laplacian pyramid techniques are utilized to augment the original landfill dataset, thereby enhancing the representation of key landfill features at various stages. Finally, we finetune the pre-trained model with the augmented dataset, further improving its capability to discern both detailed and overarching features of landfills. The experimental results demonstrate that the proposed method achieves a mean pixel accuracy (MPA) of 98.111% and a mean intersection over union (mIoU) of 88.57% in landfill segmentation, surpassing other methods in terms of accuracy and proving to be highly effective for landfill monitoring.