| The Loess Plateau accounts for 22%of the world’s apple growing area and 27%of its production,and is the world’s largest producer of quality apples.Soil moisture and soil organic carbon are important environmental factors in the nutritional and reproductive growth stages of apple trees in the Loess Plateau region,and play a vital role in the improvement of local apple yields.Therefore,real-time dynamic monitoring of soil moisture and soil organic carbon conditions in apple orchards is of great significance to the efficient management of moisture in apple orchards in the Loess Plateau region and the improvement of quality and efficiency of the apple industry.UAV multispectral remote sensing provides extremely favourable conditions for monitoring soil moisture and soil organic carbon,and has increasingly significant development advantages in practical research.At present,most of the research on soil moisture and soil organic carbon inversion using UAV multispectral remote sensing data is focused on bare soil or agricultural fields with wheat and maize as cover conditions,while relatively little research has been conducted on soil moisture and soil organic carbon in apple orchards.Therefore,this paper takes apple orchards in Yan’an City,a major apple producing area on the Loess Plateau,as the research object,and uses an UAV to obtain multispectral remote sensing image data of apple orchards at key fertility stages,and collects soil moisture content and soil organic carbon at different depths,extracts the wavelength reflectance of fruit tree canopies and calculates various vegetation indices.The model will be evaluated by using different screening methods and machine learning algorithms to construct soil moisture content and soil organic carbon inversion models for different fertility periods and different depths in apple orchards.The evaluation indexes R~2and RMSE will be used to evaluate the accuracy of the models,and the optimal inversion model for soil moisture and soil organic carbon in apple orchards will be selected to provide a scientific basis for water management and sustainable development in apple orchards.The main research findings are as follows:(1)An inverse model of soil moisture content in apple orchards was constructed for different fertility conditions at different depths.The vegetation indices constructed by the spectral bands under different fertility conditions were different,and thus showed some differences in correlation with soil moisture content.The optimal combination of variables was selected by the full subset screening method,and three machine learning methods,namely back propagation neural network(BPNN),support vector machine(SVM)and convolutional neural network(CNN),were used to construct the inverse model of soil moisture content in apple orchards under different fertility and depth conditions.The SVM model was found to be the best fit,the CNN model was the second best fit and the BPNN model achieved the worst accuracy for the same fertility period and depth conditions.The inversion of soil moisture content in apple orchards at different fertility stages showed that all three machine learning methods achieved the best results at the fruit expansion stage and poorer results at the colouring and ripening stage,but good results were achieved at different fertility stages.The SVM model performed best in the 0-40 cm depth during fruit expansion,with R~2reaching 0.911 and 0.821 and RMSE ranging from 0.003 to 0.005.(2)The inverse model of soil organic carbon in apple orchards was constructed under different depth conditions at different fertility periods.There were differences in the optimal combinations of soil organic carbon variables screened using the full-subset screening method,but the screened combinations basically passed the significance test,indicating that the full-subset screening method for screening the optimal variables of the model has certain reliability.Based on the screened optimal variable combinations,three machine learning methods,namely back propagation neural network(BPNN),support vector machine(SVM)and convolutional neural network(CNN),were used to construct soil organic carbon inversion models for apple orchards at different fertility periods and different depths.The SVM model was found to be better than the BPNN and CNN models in terms of overall fitting and prediction under the same depth conditions within the same fertility period.Comparing the machine learning models of soil organic carbon in apple orchards at different fertility stages,all three machine learning methods achieved the highest accuracy of the models at the fruit expansion stage,and the inversion of the models was poor at the colouring and ripening stage.The best monitoring depth was 0-60 cm for all fertility stages,followed by 0-40 cm and 0-20 cm.The best performance was achieved during the 0-60 cm depth of fruit expansion,with R~2above 0.61 in both the modelling and validation sets of the three models.(3)A vegetation index-based inversion model of the SOC-SMC coupling coordinated degree in apple orchards was constructed.The SOC-SMC coupling coordinated degree model showed some differences between different fertility stages,with the order of flowering and fruiting>colouring and ripening>fruit expansion,and a better coordination between soil moisture and organic carbon at 0-60 cm depth than at 0-40 cm depth within the same fertility stage.In the inversion of the SOC-SMC coupling coordinated degree,machine learning models for apple orchards at different fertility stages were compared.The SVM model outperformed the BPNN and CNN models in terms of overall fit and prediction under the same depth conditions within the same fertility period,and the model error control was relatively better than that of the BPNN and CNN models.In comparing the inversion effects of the SOC-SMC coupling coordinated degree model at different fertility periods and different profile depths,the 0-60 cm depth was better than the 0-40 cm depth at different fertility periods. |
