| Ion adsorption rare earths in southern China are widely used in high-tech fields and are important strategic resources.China has been a long-time supplier to the global rare earth market,but this has resulted in problems such as overconsumption of rare earth resources and ecological environment destruction.To protect rare earth resources and maintain sustainable development of the rare earth industry,China has entered an era of rare earth mining quota control and comprehensive mine reclamation and restoration.The precise detection of rare earth mining processes and continuous tracking of reclamation plant growth in mining areas have put forth new requirements for monitoring means.With the rapid development of high spatial resolution remote sensing monitoring technology and the field of artificial intelligence,high-precision and automatic identification of rare earth mining area feature objects has become possible.Therefore,it is important to explore new intelligent remote sensing monitoring methods to achieve informationized supervision of rare earth mining areas.In recent years,deep learning has demonstrated enhanced capability in image feature representation due to its ability to automatically extract hierarchical semantic features from images.It has achieved great success in the field of natural images.However,accurately recognizing feature objects in rare earth mining areas based on existing deep learning methods remains a challenging task.Firstly,feature objects in remote sensing scenes of mining areas are characterized by small size and dense distribution,and their features often fail to receive enough attention or become lost within deep neural networks.Secondly,environmental shadows in hilly mountainous areas and fragmentation of mining surfaces intensify semantic confusion among geographic entities.This confusion exists not only between foreground objects and background noise,but also between classes of foreground objects.In response to the aforementioned challenges,this paper takes ion adsorption rare earth mining areas in southern Jiangxi Province as the research object,and constructs an intelligent interpretation method for high-resolution remote sensing image scenes of typical objects such as rare earth leaching ponds,buildings,and reclamation vegetation based on convolutional neural networks.This method excavates the spatial semantic relationships within and between classes to construct an attention mechanism with class information constraints,which suppresses background noise and adaptively focuses on foreground target features.The main research results and contributions include the following aspects:(1)This study proposes an end-to-end framework called MCCANet for detecting mining objects and reclamation vegetation objects in rare earth mining scenarios using high spatial resolution remote sensing images.A remote sensing dataset of a rare earth mining area was constructed to validate the proposed method.The results show that the framework can achieve higher detection accuracy than the current popular methods in rare earth mining scenarios.The mining feature detection m AP value reaches 88.3%,which is 4.9% higher than the highest two-stage network,and the FPS reaches 32.1.MCCANet takes into account the ideal detection speed while maintaining high accuracy in individual object detection in the mining scenario(2)To address the interference effects caused by environmental shadows in mining scenes,a novel network head structure is constructed in this study to support multi-spectral channel information input.This introduces richer spectral dimensional features while employing a branching structure to emphasize feature differences of feature objects from the spatial dimension and spectral channel dimension,mitigating the interference of confusing visual features such as shadows.The study explores the application of multispectral channel information in remote sensing images in object detection tasks.(3)Small feature objects in mining scenes are insignificant and susceptible to complex background noise interference,exacerbating the semantic confusability among geographic entities.To address this challenge,a process of class information constraint is constructed using intra-and inter-class spatial semantic relationships.This enables the network to focus on foreground object features adaptively from the whole scene and enhances the intra-and inter-class semantic feature representation.A strong class supervision approach is imposed at the image element level to mitigate the interference effects of complex background noise in the scene and achieve more accurate discovery of potential small objects in complex mining scenes.(4)The loss function has been redesigned,and a novel multi-task loss function is constructed.This function adds a class mapping loss to the traditional loss function for object detection,strengthening the network’s ability to represent foreground objects’ features,and establishing more accurate class information constraints to optimize the model’s end-to-end training.In this paper,we combine deep learning technology and remote sensing technology to build a convolutional neural network model that is applicable to the high-performance detection of typical feature objects in the entire process from mining to reclamation in rare earth mining areas.The research provides intelligent and process-oriented supervision means for the quota control of rare earth resources and mine reclamation and restoration.This is essential for the ecological sustainable development of rare earth mining areas. |
PDF Full Text Request