Agricultural Land Use And Crop Phenological Information Extraction Based On Sentinel-2 Single Interannual Time-series Data

Crop phenology information is essential for crop identification,growth monitoring,and farmland utilization planning.The existing methods of phenological analysis usually assume that the crop rotation pattern in the study area is essentially fixed for many years,therefore crop seasons can be identified from periodic changes over a multi-year time-series vegetation indices(VI).In recent years,farmland abandonment and seasonally fallow in south-central China have become increasingly severe,and fallow farmland’s spatial and temporal distribution shows apparent randomness.In this case,setting uniform cycle parameters in the region will lead to a significant bias in the weather extraction results.Therefore,for areas with complex agricultural land use,it is necessary to establish a method that can process single inter-annual time-series VI and autonomously distinguish crop fertility to accurately extract phenological information on a large scale.An improved single interannual time-series NDVI(SI-ts NDVI)reconstruction method and an application framework for agricultural land use and phenology information extraction were proposed in this paper.The framework mainly includes(1)A Random forest(RF)classifier for identifying farmland,which was constructed with SI-ts NDVI and its Fourier harmonic features;(2)Fast Fourier Transformation(FFT)and Double Logistic(DL)functions were coupled to establish an adaptive reconstruction method FFT-L DL for SI-ts NDVI reconstruction of seasonally fallow and rotation farmland.(3)4 methods were provided to extract the waiting information from the reconstructed SI-ts NDVI.The application effect of the framework was tested and demonstrated in a study area in southcentral China.The following results were obtained:(1)An RF classifier based on SI-ts NDVI and its Fourier harmonic features was proposed to improve the accuracy of farmland classification.The SI-ts NDVI dataset was established using Sentinel-2 A/B data from May 2019 to June 2020,and 2011 sample points of 18 types of land use were selected.The harmonic features of SI-ts NDVI were calculated using FFT,then 5 RF classifier models were trained with different feature combinations,and the classifier with the highest precision was used to extract information on farmland utilization in the study area.The results indicated that the FFT harmonic features were practical for the classification of farmland utilization,with an accuracy of up to 93% for seasonally fallow,rotation,and abandoned farmlands.The framework was effective in identifying 12 types of farmland utilization and 6 types of non-agricultural land use in the study area,with an overall accuracy of 82%.(2)An adaptive reconstruction algorithm FFT-L DL was proposed to address the problem of existing reconstruction methods that cannot correctly identify the number of rotation crops from SI-ts NDVI.The algorithm first uses the FFT harmonics to locate the dates of the NDVI peak and valley.It determines the local fitting intervals,then fits the NDVI in each interval with the DL function,and finally merges the results of each interval with the global model function.The performance of the FFT-L DL,S-G filter,and Fourier filter was compared on 1136 farmland samples and the model SI-ts NDVI created from them.The results showed that FFT-L DL was not disturbed by abandoned crops and weeds and can accurately identify the number of rotating crops and achieve SI-ts NDVI adaptive reconstruction of rotation and seasonally fallow farmland.By comparing the performance on simulated data with different noise levels,FFT-L DL was significantly better than the SG and Fourier filters,and the results were more consistent with the actual crop growth state.(3)The performance for extracting SOS and EOS of Seasonal amplitude thresholding(SAT),Relative amplitude thresholding(RAT),Absolute value thresholding(AVT),and Slope extremum method(SEM)in different scenarios were compared by samples and other phenological products.The results showed that all four methods could effectively obtain phenological information,with good consistency among the results.SAT can adjust the thresholds autonomously based on each crop season and showed the most consistent performance in the experiment.SEM requires no input parameters and highly correlates with the reference phenology.Still,the identification may be unstable when the crop growth period is incomplete in SI-ts NDVI.The parameter settings of RAT and AVT significantly impact the results.The application framework proposed in this paper using SI-ts NDVI Sentinel-2 and can extract the phenological parameters accurately through a reconstruction method that autonomously determines the fertility period of rotated crops,which is suitable for application in areas with significant changes in farmland use and farming systems.It can automatically and efficiently extract crop phenological information from agricultural fields with strong adaptiveness and robustnessand,and can provide necessary data support for crop growth condition monitoring,agricultural land resource management,etc.