Soil Salinity Inversion Model Based On UAV Multispectral Remote Sensing

Soil salinization seriously threat to the high-quality development of agriculture in the Hexi oasis irrigation areas in China.Rapid and accurate understanding of the distribution of saline-alkali land within the farms in the irrigation areas is crucial for effective prevention and control of saline-alkali land and ensuring food security in China.In view of this,this study takes typical saline and alkaline land in the Hexi oasis region as the research object,obtains multispectral remote sensing images of film mulched farmland,bare land,alfalfa mulched farmland,and wheat mulched farmland in the study area through unmanned aerial vehicles,constructs soil salt inversion models based on machine learning algorithms,and conducts precision evaluation to optimize the optimal inversion model for soil salt,It is expected to provide theoretical basis for accurate fitting of soil salinity inversion models under different crop mulches.The main contents are as follows:(1)In order to find out the influence of film mulching on the accuracy of soil salinity retrieval in saline alkali cultivated land,a soil salinity retrieval model was constructed based on support vector machine(SVR),backpropagation neural network(BPNN)and random forest(RF)algorithm.The results show that the inversion model after removing the background of film mulching has better inversion performance than the model before removing the background of film mulching.The use of neural network classification method to remove the background of film mulching can effectively improve the accuracy of soil salt inversion in film mulching farmland.The accuracy of soil salt inversion model based on salt index is higher than that based on spectral reflectance.The salt index as the input layer of the model can improve the accuracy of the model.Compared with three soil salinity retrieval models constructed by the salinity index,the precision of the random forest retrieval model after removing film mulching is the best,the determination coefficient of the validation set is 0.743,and the root mean square error is 0.091.Support vector machine models have the second highest accuracy,while backpropagation neural network models have relatively poor performance.(2)To improve the accuracy of the backpropagation neural network(BPNN)model in inverting soil salinity in cultivated land,this study used particle swarm optimization(PSO),thought evolution algorithm(MEA),and genetic algorithm(GA)to optimize the BP neural network soil salinity inversion model and determine the optimal optimization algorithm.The results show that the PSOBPNN,MEA-BPNN,and GA-BPNN models can effectively reflect soil salt content.Optimizing the BPNN model with PSO,MEA,and GA algorithms can effectively improve the accuracy of BPNN model in soil salt inversion.The GABPNN model is the optimal soil salt inversion model.Compared to the MEABPNN model and PSO-BPNN model,the GA-BPNN model has a coefficient of determination increase of 0.133 and 0.256,a root mean square error decrease of 0.034 and 0.037,and a performance deviation ratio increase of 0.230 and 0.478.The GA algorithm has better optimization effects on the BP neural network model than the MEA algorithm and PSO algorithm.(3)In order to improve the accuracy of the BPNN soil salt inversion model,this study constructed an optimized BPNN soil salt inversion model for mulched farmland based on spectral reflectance and salt index.The results showed that the accuracy of the model with salt index as the input layer was significantly higher than that of the soil salt inversion model based on spectral reflectance.The GA-BPNN model had the best accuracy improvement,and the validation set determination coefficient Rv2 after removing the mulched background increased by 0.193,The root mean square error RMSEv decreased by 0.094,followed by the MEA-BPNN model and the PSO-BPNN model.(4)In order to build soil salt retrieval models under different mulching conditions,this study built soil salt retrieval models for bare land,alfalfa covered land and wheat covered land as surface crops based on support vector machine(SVR),back propagation neural network(BPNN),random forest(RF)algorithm and extreme learning machine(ELM)algorithm.The results show that the improved spectral index calculated by introducing the red edge band can improve the accuracy of soil salt inversion under crop cover.For bare land,the best model in the soil salt inversion model established using the salt index is the ELM model,with a validation set determination coefficient of 0.707,a root mean square error of 0.290,and a performance deviation ratio of 1.852;The BPNN model is the best model for establishing soil salinity inversion models using vegetation index for agricultural farmland with vegetation coverage.The validation set determination coefficient reached 0.836,the root mean square error was 0.027,and the performance deviation ratio was 2.100.