Att-Unet: An Improved Moho Inversion Appoach from Remote Sensing Data

Density interface inversion is vital for determining the relief and depth variations. Deep learning algorithms, using gravity data, can efficiently generate inversion results. However, gravity inversion problems often suffer from ill-posedness, causing non-unique solutions. While most deep learning methods approximate density interface trends, they lack sensitivity for fine details and generalize poorly in complex terrains. This study proposes an improved Att-Unet model, inspired by the Attention Gate algorithm, to recover depths and relief from gravity anomaly data. Additionally, the loss function is enhanced to emphasize variations in small target reliefs. The dataset is created by randomly assembling, superimposing, and smoothing rectangular models. Using the corrected Parker forward theory, the training dataset for output interfaces is generated. The method is validated by comparing inversion results with U-net and Parker-Oldenburg methods, using real data from Brittany, France. Real experiments show the model’s sensitivity and accuracy, reducing the RMS difference from 0.3168 to 0.1879.