Recognition Of Heavy Metal Stress In Rice Based On Remote Sensing Phenology And Data Mining Algorithms

Heavy metal pollution in soil can cause stress on crop growth,which results in increasingly serious food security.Therefore,Methods for rapidly and accurately monitoring heavy metal stress have important practical significance.Remote sensing,with its objective,real-time,dynamic and multi-spectral characteristics,has become a powerful tool for heavy metal stress research in crops.On the basis of previous studies,we studied heavy metal stress in rice using machine learning from the perspective of remote sensing phenology.The rice fields polluted by heavy metals in Zhuzhou,Hunan province were selected as the experimental areas.The data collected include leaf area index(LAI),heavy metal concentration and meteorological data at several critical rice growing stages,and remote sensing data,including CCD images of HJ-1,ETM+ images of Landsat 7,OLI images of Landsat 8,and MOD09A1 product.The identification method of heavy metal stress in rice was explored based on remote sensing phenology and data mining algorithms,and the classification of heavy metal stress levels in rice in the study area was realized.The major work and conclusions are as follows:(1)Combined with agronomic rules and derivation operation,the key phenological periods of rice were extracted from the remote sensing vegetation indices time-series,and the results indicate that heavy metal poisoning can lead to pheonological changes,and the phenological differences existed under different heavy metal stress levels.In order to solve the problem of low accuracy in identifying heavy metal stress level by single phenological indicator,a new way for establishing remote sensing phenological index was designed,which was to mine more in-depth phenological information in time series data and extract more sensitive phenological indicators to heavy metal stress,improving the recognition accuracy of heavy metal stress levels.(2)The rice root system is the part that directly contact with heavy metals in soil,and it is more poisoned by heavy metals than other organs such as stems and leaves.In this paper,the phenological parameters in the WOFOST model were set up according to the key phenological period extracted,and the images that used for the assimilation were determined.The assimilation framework of the RS-WOFOST was run to obtain the change of the dry weight of roots.It was found that the dry weight of roots under different stress varied greatly at heading stage,by comparing the curves of dry weight of roots under different heavy metal stress.Therefore,the dry weight of roots at heading stage was selected as a sensitive biological feature for identifying heavy metal stress.(3)The phenological indicators and WRT values obtained from the above studies were taken as high-dimensional feature set for remote sensing identification of heavy metal stress,and a strategy for feature optimization was also designed.The results indicate that compared with the original high-dimensional feature set,the classification accuracy for heavy metal stress levels can be improved by 1% based on the optimal feature subset.(4)According to the idea of ensemble learning,random forest(RF)and gradient lifting(GB)algorithms were integrated,and an ensemble classification model,which based on the optimal feature subset,was established to classify the heavy metal stress levels.The results show that the ensemble model can achieve accurate classification of stress levels,and the overall accuracy of discrimination for different stress levels is greater than 98%,which increased by 2%-3% compared with that using RF or GB algorithm alone.