Abstract: To improve the prediction accuracy of phosphorus content in black soil, a prediction method based on an improved deep neural network (DNN) was proposed using thermal infrared (TIR) images. First, unmanned aerial vehicles equipped with TIR sensors were used to collect TIR images of black soil, and preprocessing such as geometric correction and atmospheric correction was conducted.Second, to address the difficulty of global optimization of DNN hyperparameters, the advantages of the bacterial foraging optimization algorithm (BFOA) and particle swarm optimization (PSO) were integrated to optimize the number of hidden layers of the DNN model. Finally, the improved DNN model was used to predict the phosphorus content in the TIR images of black soil, and the model performance was validated with measured data. The results showed that the number of hidden layers of the improved DNN model (DNN-BFOA-PSO) was 3. The root mean square error (RMSE) and mean absolute error (MAE) of the improved DNN model predictions were 1.08% and 1.42%, respectively, and the coefficient of determination (R²) was 0.99. Compared with the baseline DNN model, the RMSE and MAE decreased by 88.4% and 85.6%, respectively, and R² increased by 0.09. Compared with random forest, feature selection combined with random forest, and partial least squares regression models, the improved DNN model achieved the best prediction performance for phosphorus content in black soil.

Key words: deep neural network (DNN), black soil, phosphorus content prediction, BFOA algorithm, PSO algorithm