Aiming at the time-consuming and labor-intensive problem of traditional soil bulk density measurement of topsoil, using easily available soil physical parameters to accurately and quickly predict the bulk density of topsoil in farmland. By analyzing the relationship between soil bulk density of topsoil layer and surface roughness and resistance of soil, a gradient boosting regression tree (GBRT) model with input of surface roughness and resistance of soil was constructed. The roughness of soil surface was obtained using image processing techniques. Using homomorphic filtering technology to preprocess the surface image of soil, extract the entropy, average, variance, skewness and kurtosis of the image gray histogram to characterize the texture feature parameters of image, extract the energy, entropy, contrast and inverse variance characterize the regional characteristic parameters of the image. The soil resistance was obtained using a laboratory vehicle-mounted resistance measurement system. Using gray correlation analysis, from nine characteristic parameters that characterizing the roughness of soil surface and soil resistance, the variables with bulk density of topsoil greater than 0.65 were selected as the model input. The prediction results of the GBRT model were the same as those obtained by the ring knife method. As a result of correlation analysis, the determination coefficient R2 reached 0.8782, and the average absolute error reached 0.021g/cm3. At the same time, under the same input parameters and computing environment, compared with the prediction accuracy and operation speed of the BPNN and SVR models, it was verified that the GBRT model had better prediction accuracy and shorter operation time. The research results can provide ideas for obtaining the bulk density of topsoil and provide theoretical support for scientific and rapid guidance of farmland.