Search

Remote sensing estimation of maize residue cover (MRC) can rapidly acquire large-scale data on MRC, crucial for monitoring and promoting conservation tillage. Herein, seven tillage indices derived from Sentinel-2 satellite imagery were analyzed alongside measured MRC data to assess their correlation with MRC. MRC estimation models were built using six machine learning algorithms, including back propagation neural network (BPNN), random forest (RF), support vector regression (SVR), extreme gradient boosting (XGBoost), Stacking1, and Stacking2. Model performance was compared using the coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). The potential for conservation tillage was explored. The results showed that the R2 values of the seven tillage indices in the study area exceeded 0.5, with particularly high correlations for NDTI and STI, with R2 values of 0.755 and 0.751, respectively. When using machine learning algorithms to construct models, the Stacking2 model exhibited the highest estimation accuracy, with an R2 of 0.923, RMSE of 3.32%, and MAE of 0.025, while Stacking1 also demonstrated robust performance, with an R2 of 0.910, RMSE of 3.45%, and MAE of 0.029. Among the base models, XGBoost achieved the highest estimation performance and the lowest error, with R2, RMSE, and MAE values of 0.884, 4.77%, and 0.031, respectively. The R2 values of RF, SVR, and BPNN were 0.865, 0.859, and 0.842, respectively, with RMSE values of 4.06%, 4.76%, and 5.91%, and MAE values of 0.039, 0.047, and 0.059, respectively. These results indicate that the Stacking2 model demonstrates a significant advantage in prediction accuracy. Geostatistical analysis of the inversion results of the Stacking2 model revealed that the proportions of farmland with MRC values exceeding 30% in Changchun, Songyuan, and Siping were 81.90%, 77.96%, and 83.58%, respectively. This indicates that Changchun and Siping have greater potential for implementing conservation tillage. This study demonstrates that the stacking ensemble learning model significantly improves the predictive performance by leveraging the strengths of multiple base models and accurately monitoring the spatial distribution of MRC.

The management of crop residues in farmland is crucial for increasing soil organic matter and reducing soil erosion. Identifying the regional extent of crop residue cover (CRC) is vital for implementing conservation tillage and formulating agricultural subsidy policies. The Google Earth Engine (GEE) and remote sensing images from 2019 to 2023 were used to obtain spectral characteristics before the maize seedling stage in Northeast China, followed by constructing the CRC estimation models using machine learning algorithms. To avoid the impact of multicollinearity among data, three machine learning algorithms—ridge regression (RR), partial least squares regression (PLSR), and least absolute shrinkage and selection operator (LASSO)—were employed. By comparing the accuracy of these methods, the most accurate model was determined and applied to subsequent CRC estimation. Based on the estimated CRC and Conservation Technology Information Center definitions of tillage practices, the conservation tillage mapping was completed, and the spatiotemporal distribution characteristics were thoroughly analyzed. The following findings were demonstrated: (1) the PLSR-based model outperformed RR (Pearson’s correlation coefficient (r) = 0.8875, R2 = 0.7877, RMSE = 6.99%) and LASSO (r = 0.8903, R2 = 0.7926, RMSE = 6.88%) with higher accuracy (r = 0.9264, R2 = 0.8582, RMSE = 4.93%). (2) Over the five years, the average no-tillage (NT) proportion in the study area was 15.9%, reduced tillage (RT) was 17.8%, and conventional tillage (CT) was 66.3%. In 2020 and 2022, NT rates were significantly higher at 27.5% and 15.5%, while RT were 15.7% and 30.0%, respectively. (3) Compared to the Sanjiang and Liaohe Plains (RT = 1907 km2 and 1336 km2, and NT = 559 km2 and 585 km2, respectively), the Songnen Plain exhibited higher conservation tillage rates (where RT was 3791 km2 and NT was 1265 km2). This provides crucial scientific evidence for the management and planning of conservation tillage, thereby optimizing farmland production planning, enhancing production efficiency, and promoting the development of sustainable agricultural production systems.

The spectral features of soils are a comprehensive representation of their physicochemical parameters, surface states, and internal structures. To date, spectral measurements have been mostly performed for powdered soils and smooth aggregate soils, but rarely for cracked soils; a common state of soda-saline soils. In this study, we measured the spectral features of 57 saline soil samples in powdered, aggregate, and cracked states for comparison. We then explored in depth the factors governing soil spectral features to build up simple and multiple linear regression models between the spectral features and physicochemical parameters (salt content, Na+, pH, and electronic conductivity (EC)) of saline soils in different states. We randomly selected 40 samples to construct the models, and used the remaining 17 samples for validation. Our results indicated that the regression models worked more effectively in predicting physicochemical parameters for cracked soils than for other soils. Subsequently, the crack ratio (CR) was introduced into the regression models to modify the spectra of soils in powdered and aggregate states. The accuracy of prediction was improved, evidenced by a 2⁻11% decrease in the parameters mean absolute error (MAE).