计算机科学
遥感
合成孔径雷达
人工智能
深度学习
保险丝(电气)
传感器融合
分割
混淆矩阵
散斑噪声
计算机视觉
高光谱成像
多光谱图像
判别式
GSM演进的增强数据速率
融合
噪音(视频)
模式识别(心理学)
城市固体废物
斑点图案
冗余(工程)
环境科学
残余物
图像融合
数据挖掘
编码器
图像分辨率
混乱
自编码
偏移量(计算机科学)
雷达成像
情态动词
图像分割
卷积神经网络
作者
Jiahua Teng,Songjing Guo,Wenfu Wu,Jiaqi Yu,Yunhao Chen
摘要
ABSTRACT Municipal solid waste (MSW), a global environmental concern whose monitoring is critical for environmental governance, can be effectively detected by remote sensing. Optical remote sensing is disturbed by weather, illumination, and vegetation, whereas synthetic aperture radar (SAR) data suffers from speckle noise and poor semantic characterization. Coupled with material heterogeneity, fuzzy boundaries, and spectral confusion in open‐pit MSW dumps, current single‐modal models cannot achieve accurate pixel‐level detection, hindering large‐scale automatic MSW mapping. Therefore, this study proposes MSWNet, a dual‐modal fusion deep network for high‐precision MSW detection, and builds the first pixel‐labeled MSW dataset of co‐registered optical‐SAR imagery. Specifically, to mitigate spectral confusion and material heterogeneity within MSW, a correlation‐guided cross‐modal fusion module is designed to adaptively weight optical‐SAR features, eliminate modal redundancy, and fuse discriminative multidimensional features for MSW‐background discrimination. Meanwhile, RefineNet, a lightweight residual edge refinement subnetwork deployed after the decoder, extracts edge details from upsampled coarse detection outputs to offset encoder downsampling‐induced edge loss and correct boundary misclassification for sharper MSW detection edges. Quantitative experiments demonstrate that the proposed MSWNet outperforms eight mainstream segmentation models with state‐of‐the‐art metrics (OA = 98.41%, Kappa = 0.9234, mIoU = 0.9277). This study can provide reliable technical references and fine spatial data for monitoring and intelligent management of open‐pit MSW.
科研通智能强力驱动
Strongly Powered by AbleSci AI