Machine Learning-based Remote Sensing Inversion Study Of Key Crop Parameters

Crop chlorophyll and biomass inversion is of great significance and value,which can not only provide data support for improving crop production efficiency and quality,but also predict crop yield,protect agroecological environment,and promote sustainable agricultural development.Therefore,strengthening the research and application of crop chlorophyll and biomass inversion techniques will have a positive impact on agricultural production and sustainable development.Both traditional chlorophyll content measurement methods and traditional biomass monitoring methods have many drawbacks,such as the time and labour consuming nature of sample plot surveys,the random nature of field surveys,the small area covered,and the susceptibility to different topographical and land use factors.In view of this,this experiment is divided into two parts:1.The PROSALI radiative transfer model was used to build a crop canopy reflectance-canopy chlorophyll training dataset for the inversion of chlorophyll over a regional area,and combined with Sentinel-2 satellite remote sensing data for the quantitative inversion of crop chlorophyll in the Hailun region.The training dataset was divided into two parts: PROSAIL-training dataset and field collection-training dataset,which were used to train the artificial neural network model and the BP neural network model respectively.The trained neural network models were then used to invert the canopy chlorophyll content of the crops.The results show that the results of the PROSALI-based remote sensing inversion of chlorophyll in crops are in good agreement with the measured chlorophyll,and the accuracy of the inversion of chlorophyll concentration is significantly improved compared with the inversion results of the training data set constructed with field collection.2.The inversion of above-ground biomass in a small-scale area was done with soybean as the study object,acquiring UAV multispectral images of soybean in different periods and at different heights.Eight vegetation indices were constructed by reconstructing five bands of the UAV for the key periods of soybean growth,podding and bulging,and doing a correlation analysis of the UAV images at different resolutions at seven heights combined with ground collection biomass.Eight vegetation indices with strong correlation with biomass were established to build regression models,and three machine learning methods,namely random forest,support vector machine and 1D deep learning neural network,were used to construct biomass inversion models for different time periods and different resolutions of soybean,respectively.The results showed that:(1)In the 1D deep learning neural network model: the R2 reached 0.74 and 0.71 at 50 m and 0.78 and 0.72 at 70 m.(2)The accuracy of the inversion was better at the podding stage than at the bulge stage due to the flight weather and the soybean growing period.(3)The inversion accuracy was best when the UAV was flown at an altitude of 70 m.