Gravity anomaly inversion is a technique used to estimate underground density distribution using gravity data. This paper proposes a new three-dimensional (3D) gravity anomaly inversion method based on the LinkNet network. Compared with two-dimensional gravity anomaly inversion, 3D gravity anomaly inversion can determine the density distribution of the entire region below the observation surface. Additionally, compared with traditional methods, the neural network method does not require the selection of initial parameters, and several predictive models can be quickly sought during the prediction stage. The Tversky loss was used to improve the inversion accuracy of the boundary. By comparing the inversion of the fully convolutional network, UNet network, and LinkNet network proposed in this paper on simulated data, it was observed that the model reconstruction error obtained using the LinkNet network had the best fitting effect with gravity data, which were 0.3526 and 0.0521. The results reveal that this method can achieve accurate inversion. Using the San Nicolas deposit in Mexico as an example, the proposed method and the improved preconditioned conjugate gradient algorithm were compared to further illustrate the effectiveness of the algorithm. The results reveal that the position and shape trends of the geological body attained using the proposed approach are in good agreement with the drilling data.

重力异常反演是一种利用重力数据估计地下密度分布的技术。本文提出一种基于LinkNet网络的三维（3D）重力异常反演新方法。与二维重力异常反演相比，三维重力异常反演可以确定观测面以下整个区域的密度分布。另外，与传统方法相比，神经网络方法不需要选择初始参数，并且在预测阶段可以快速寻找多个预测模型。采用Tversky损失来提高边界反演精度。通过在模拟数据上对比本文提出的全卷积网络、UNet网络、LinkNet网络的反演，结果发现，利用LinkNet网络得到的模型重建误差与重力数据的拟合效果最好，分别为0.3526和0.0521。结果表明该方法能够实现准确反演。以墨西哥圣尼古拉斯矿床为例，将所提方法与改进的预处理共轭梯度算法进行比较，进一步说明算法的有效性。结果表明，利用该方法获得的地质体位置和形状趋势与钻探数据吻合较好。将所提出的方法与改进的预条件共轭梯度算法进行比较，进一步说明该算法的有效性。结果表明，利用该方法获得的地质体位置和形状趋势与钻探数据吻合较好。将所提出的方法与改进的预条件共轭梯度算法进行比较，进一步说明该算法的有效性。结果表明，利用该方法获得的地质体位置和形状趋势与钻探数据吻合较好。