Inversion and mapping of soil alkali-hydrolyzable nitrogen in farmland using satellite remote sensing and machine learning

Abstract

Soil alkali-hydrolyzable nitrogen (SAN) is a crucial indicator of soil nutrient status. Quickly and accurately obtaining the spatial distribution of SAN content is essential for the effective implementation of precision variable fertilization. To efficiently estimate SAN content at the field scale and improve the model's accuracy and stability, the basic farmland of Jianshan Farm in Heilongjiang Province was selected as the study area. Remote sensing factors and topographic variables were extracted from Sentinel-2 satellite imagery and Copernicus digital elevation model. Simultaneously, SAN content was measured during different bare soil periods in spring and autumn. This study employed support vector regression (SVR), extreme gradient boosting (XGBoost) and BP neural network (BPNN) to establish SAN content inversion models for different bare soil periods. Then, the zebra optimization algorithm (ZOA) was used to optimize the hyperparameters of the best-performing model, and the experimental results were analyzed. Finally, the model's interpretability and reliability were validated through feature importance analysis and statistical significance tests. The results indicated that the correlation between the raw bands of remote sensing data in autumn and SAN content was generally higher than that in spring. The spring SAN inversion model, ZOA-XGBoost-2, which incorporates raw bands and topographic variables demonstrated the best predictive performance (R²=0.3416, r = 0.6153, RMSE=31.1616 mg/kg, MAE=21.2132 mg/kg, MAPE=7.5184 %). The autumn SAN inversion model, ZOA-BPNN-6, based on raw bands, vegetation indices and topographic variables demonstrated the best predictive performance (R²=0.7304, r = 0.8664, RMSE=20.3511 mg/kg, MAE=15.5451 mg/kg, MAPE=5.2008 %). Compared with the unoptimized inversion model, R² increased by 6.32 percentage points and 1.12 percentage points, and r increased by 3.99 percentage points and 1.01 percentage points, respectively, while RMSE, MAE and MAPE decreased. In the spring model, vegetation indices and raw bands accounted for 52.82 % and 47.18 % of the total importance, respectively, while in the autumn model, raw bands, vegetation indices, and topographic variables contributed 46.15 %, 30.72 %, and 23.13 %, respectively. In addition, ANOVA confirmed that the model performance differed significantly between the two seasons, validating the superiority of the autumn model in SAN prediction. The proposed inversion model yielded better prediction results, which closely aligned with the actual SAN data. It can provide a theoretical basis and technical support for the rapid and dynamic monitoring of soil nitrogen content at the farm scale, as well as for variable fertilization decision-making.