Objective The aim was to explore the gamma spectrum data to predict the spatial distribution of fertility in the soil surface layer (0 ~ 30 cm).

Method The soil fertility factor models were established by using partial least squares regression (PLSR), support vector machine (SVM) and BP neural network (BPNN).

Result The soil fertility factors prediction accuracy of BP neural network model was better than partial least squares regression and support vector machine. The prediction accuracy of BP neural network model for total nitrogen content, pH, clay content and sand content were higher, R² values were 0.564, 0.556, 0.612 and 0.626, respectively. However, the prediction accuracies for total potassium and total phosphorus contents did not meet expectations. Compared with the actual distribution results of sample points, the prediction results of BP neural network of soil total nitrogen content, pH, clay content and sand content in the study area were basically consistent in numerical statistical characteristics and trends.

Conclusion It was feasible to predict the spatial distribution of soil total nitrogen content, pH, clay content, and sand content using gamma spectroscopy data in the research area, but total phosphorus, total potassium and silt content could not be predicted.