Research On Rice Extraction In Large Areas With Time-series Remote Sensing Images Considering Feature Screening

Rice is one of the three main staple foods in China,and ensuring the cultivation of rice is closely related to national food security.Accurately obtaining the spatial distribution and planting area of rice is of great practical significance for the government to adjust the agricultural industrial structure,timely formulate targeted policies,and achieve agricultural power and sustainable economic development.In recent years,the rapid development of remote sensing technology has provided unprecedented opportunities for macroscopic monitoring of rice distribution information,especially with the increasing data sources and the improvement of image time and spatial resolution,which provide a good data basis for precise extraction of rice.However,achieving largearea regional rice monitoring and extraction requires downloading,processing,and analyzing large amounts of data.Further exploration and research are needed to utilize the phenology and phenological characteristics of rice.The emergence of remote sensing big data cloud platforms provides strong support for rapidly extracting the planting range of rice in large areas.This thesis presents a study on the time-series extraction of rice in Heilongjiang Province in 2020 and 2021 using multi-source remote sensing data such as Sentinel-2,Sentinel-1,and SRTM on the Google Earth Engine cloud platform.By selecting and analyzing the characteristics of rice during multiple phenological stages,various machine learning methods were used to conduct rice extraction research from both single and multi-temporal perspectives,with or without feature selection.The study obtained highly accurate results for rice extraction in Heilongjiang Province.The main research content and conclusions of this article are as follows:(1)Analysis and selection of image characteristics at different phenological stages:We constructed multiple features from various remote sensing data sources and optimized and selected feature sets based on J-M distance according to the separability size of rice and other ground objects during different phenological stages.After optimization,37 features were screened down to 11 features,including MNDWI,GCVI,B11,elevation,B8＿asm,and VH.The analysis results indicated that the optimized features can better separate rice from other ground objects during the corresponding phenological stages.(2)Extraction of rice based on different phenological stages: This study conducted single-phenology period-based and joint-multi-phenology period-based extraction studies to explore the optimal phenological period for extracting rice in Heilongjiang Province.According to the growth and development timepoints of rice and the availability of remote sensing time-series images,we divided them into four phenological stages: bare soil stage,planting stage,growth stage,and harvesting stage.During the planting stage,six types of features including MNDWI,GCVI,B11,elevation,B8＿asm,and VH were selected and used for rice extraction,achieving an overall accuracy of 96.9%,higher than the highest overall accuracy of 88.4% during the growth stage and 88.1% during the harvesting phase.This indicates that the planting period is the best single-time frame phenological period for accurate extraction of rice.By selecting eleven optimized features for joint multi-phenology period-based time-series rice extraction,the classification results showed a higher overall accuracy of 97.5%.Furthermore,compared with the nonfeature-selected results,the optimized features reduced the frequency of misclassification of non-rice spectra.(3)Rice extraction based on different classification methods: This study compared the results of rice extraction using three machine learning algorithms-classification regression trees,support vector machines,and random forests-for single-time frame and multi-temporal angle-based rice extraction experiments.Random forest achieved higher precision in rice extraction than classification regression trees and support vector machines.Moreover,its classifier parameter adjustment range is more convenient and reasonable,making it more suitable as an algorithm for extracting large-area rice.The statistical analysis of the rice extraction results showed a high correlation with the rice planting area of Heilongjiang Province,with a linear regression correlation coefficient of0.992.In conclusion,this thesis extracted the rice planting area in Heilongjiang Province through remote sensing interpretation and proposed a new approach based on feature optimization and selection during different phenological stages to enhance the temporal information of classification pixels.The high-precision extraction results of rice planting range were obtained.In future research,combining higher resolution image data with deep learning algorithms will lead to new breakthroughs in large-area rice remote sensing extraction.The technical route and research idea of this study provide a reference for remote sensing extraction of rice in large areas...