Background Regional soil erosion spatial distribution information plays an important role in ecological restoration and land use optimization decision-making, but the spatial model for its analysis and calculation is not yet mature. Deep learning method was introduced in order to establish the complex relationship between soil erosion and various factors and to obtain high-precision spatial distribution data of soil erosion intensity by using its strong computing ability and good fitting effect. The large area of coniferous forest cover in Hubei leads to serious soil erosion, which was taken as research area to verify the feasibility and efficiency of deep learning in regional soil erosion spatial distribution information acquisition.

Methods This study introduced machine deep learning method to explore a new way to study the spatial distribution of soil erosion. The framework of UNet++ and BP neural network were constructed, and hyper-parameters were optimized on the Jupyter Notebook platform. Based on the real spatial distribution data of soil erosion in Hubei province, optimization function and loss function were used to train neurons to record the deep information of soil erosion factors. Spatial distribution data of factors were obtained by remote sensing as model input, pixel Windows were randomly extracted as training samples to calculate the spatial distribution data of soil erosion intensity grade.

Results The results showed that the overall accuracy of the UNet++ neural network model was 95.7%, and that of the BP neural network model was 91.4%. The UNet ++ model achieved the better overall accuracy than BP neural network model. The results of BP neural network model had more "salt and pepper" phenomenon, while the results of UNet++ neural network model were difficult to find. UNet++ neural network model overcame the phenomenon of "pepper and salt" in BP model. The error distribution of UNet ++ model in each erosion intensity was relatively uniform, without obvious error aggregation phenomenon. Compared with the BP model, UNet ++ model better reflected the distribution of soil erosion.

Conclusions It is proved that when rainfall erosivity, soil erodibility, land cover, vegetation cover, slope and topographic relief are selected as input factors, deep learning model can be used to automatically obtain spatial distribution data of soil erosion intensity accurately and quickly by computer. In addition, compared with the traditional BP neural network, the spatial distribution results of soil erosion intensity obtained by UNet++ model have higher accuracy and better effect.