Rapidly acquiring disturbed patches in production and construction projects through high-resolution remote sensing images holds significant importance for enhancing soil and water conservation supervision capabilities and controlling human-induced soil erosion. Traditional visual interpretation methods for identifying disturbed patches require substantial effort and time, leading to numerous limitations. To improve the efficiency of soil and water conservation supervision, this paper analyzes and summarizes the change characteristics of production and construction projects between two periods of remote sensing images. An intelligent extraction method for disturbed patches in these projects is proposed, based on deep learning and high-resolution remote sensing images. The U-Net++ architecture is employed as sub-network to construct a Siamese network model, with the integration of an attention mechanism module to enhance model performance. Experimental results in the validation area demonstrate that the proposed method achieves a detection rate of 91.52% for disturbed patches, with a false-negative rate of 8.48%. This outperforms the disturbance patch detection rate of 87.28% and a false-negative rate of 12.72% achieved by the dual-temporal early fusion strategy. The extracted boundaries of disturbed patches closely align with manually annotated patch boundaries, indicating the feasibility of utilizing deep learning for extracting disturbed patches in production and construction projects. This approach offers a novel perspective to enhance the efficiency of soil and water conservation supervision.
|