Research On Remote Sensing Monitoring Of Crop Types In Early Growing Season Combined With Planting History Information

The spatial distribution information of crop types is an important basic data for agricultural refinement management.In the current crop identification research based on remote sensing technology,many methods use the variation characteristics of spectral time series data of crops in the growing season to distinguish different crop types,and have achieved good post-season classification results.However,there is often a lack of effective time series data in the early growing season of crops,which makes it difficult for this method to achieve better identification results in the early growing season.Recently,some studies have achieved certain results in the extraction of crop planting area in the early growing season based on the phenological differences of different crops in the growing season,or the change law of crop planting in the survey statistics.Regional crop planting types often show a certain regularity in the inter-annual variation,but they have not been fully utilized in the current crop classification research.In view of the above research status of early growing season crop identification,based on deep learning technologies such as 1D-CNN and LSTM,a method for early growing season crop identification combined with planting history information is developed.In order to test the validity and reliability of this method for crop identification in the early growing season,a number of exploration experiments were carried out in different research areas such as agricultural planting areas in the United States and Sichuan in China.The main research work and achievements of this thesis are as follows:The study is based on three different spectral time series data from HLS and multi-year CDL data to conduct early crop type monitoring and identification experiments in the US study area.The experimental results show that the method of combining planting history information in areas with different planting rules can significantly improve the classification accuracy of single spectral data in the early growing season,and the early classification accuracy of S30 time series data products with higher temporal resolution in the original image has been improved more significantly.The experimental results based on planting history data of different lengths show that: the length of the planting history sequence and the accuracy of crop identification are not positively correlated.In areas with large annual changes in planting types,the planting data in recent years close to the forecast year have achieved better identification results.And for the same crop,the model based on planting history data has the highest F1-Score coefficient for the identification results of the crop planting frequency at the median and the highest value.The experimental results of adding random errors to the planting historical data show that: the noise resistance performance of the models based on different data is not consistent.The fusion model is more susceptible to random errors in the early growth period.However,in most cases,the crop identification ability after combining with planting history information is still better than that of spectral data.The Sichuan study area lacks years of available planting history data,and the spectral time series is also more susceptible to basin clouds.In order to test the effectiveness of the fusion model in the above situation,this study produces the2014-2020 crop type layers based on the Google Earth Engine（GEE）platform and the Landsat observation time series data in previous years,and combines the Sentinel time series data in the 2021 crop growing season to complete early crop type monitoring and identification experiments in this region.The experimental results show that compared with the spectral data,the Kappa index of the classification results of the fusion model in the early growing season is increased by nearly 30 percentage points and can reach the value of recognition accuracy of single spectral data in the late growing season 40 days earlier,which proves that the fusion model can effectively improve the early crop identification effect in the cloudy and foggy areas lacking planting calendar data such as Sichuan.