Research On Methods Of Crop Identification Using MODIS Time Series Data

The spatial distribution information of crop type is an important basic data for crop growth monitoring,crop yield estimation,crop planting pattern adjustment and optimization,agricultural climate and ecology modelling.Since each crop type has its specific growing pattern and phenological characteristic,how to comprehensively utilize such crop seasonal dynamic characteristics has become the key factor to distinguish the specific crop from other crops and green vegetation.Due to the wide swath and rich spectral and temporal information,MODIS data has been demonstrated to be superior in describing crop phenological characteristics and mapping crop at large scale.However,the inherent coarse spatial resolution makes MODIS suffer from the“mixed pixel”problem and consequently limit the classification accuracy.Therefore,how to make full use of the superiority in spectral and temporal information and make up for weakness in spatial resolution for MODIS data so as to effectively identify multiple crop types at large scale is a key scientific issue for current agricultural monitoring.In this study,an important grain production region in china,Heilongjiang Province,was selected as the study area where the main crop types?i.e.,corn,rice,soybeans and wheat?were the targeted land cover classes in this study.In addition,time-series MODIS data,crop reference maps of high spatial resolution and agricultural census data were employed as the main study datasets to explore the methods for accurate crop mapping.Main conclusions were drawn from this study as follows:?1?Evaluations on impacts of spectral and temporal features on crop identification based on remote sensing images.Ten feature subsets composed of different number of spectral and temporal features were extracted based on MOD09A1 time-series data and then were classified using the support vector machine method,and finally their classification accuracies were assessed.Results show spectral diversity and time variety are both vital factors for crop identification.Spectral and temporal information can be complementary for each other to provide the necessary information to identify crops.Increasing temporal resolution is not necessarily important for improving the classification accuracies for crops.The images associated with key phenological phrases can achieve relatively high classification accuracy even with a few images.?2?Evaluations on the potential of extension ways of global pairwise seperability index?SI?in feature selections for crop identification.Two approaches of extending pairwise SI to global measures were proposed,i.e.,“Average”(SI_ave)and“Minimum”(SI_min).Feature interpretability and classification accuracy of different crops were evaluated for both two approaches.The SI_ave approach generally has relatively high feature interpretability due to its better description of crucial phenological characteristics of different crops.However,those crops with low separability are sensitive to the extension approach and have higher classification accuracies by the SI_min approach.Due to higher temporal autocorrelation,the optimal features for crop classification that are selected by the SI_ave approach exhibit greater information redundancy across the time domain,resulting in relatively lower classification accuracy.?3?A study on Spectro-Temporal Automated Feature Selection?STAFS?method.STAFS can automatically select the optimal spectral-temporal features for crop identifications by taking account of feature separability and information redundancy.Thirty-four optimal features were selected for corn in Heilongjiang Province from 155 candidate features based on STAFS method.Three leaf and milky mature were found to be the most crucial periods in separating corn from other land cover classes.The best features selected by STAFS combined with the support vector machine classifier can generate high accuracies of spatial distribution and area estimation for crop types.Additionally,STAFS has high robustness and can be generalized to any other land cover types,study areas and remote sensing images.?4?A study on the method of fusing remote sensing images and census data to extract the sub-pixel crop fractions.A random forest regression model combined with a“backward elimination”feature selection strategy was used to generate the sub-pixel crop fraction with the highest accuracy.An“Iterative Area Gap Allocation”?IAGA?approach was proposed to spatially allocate the area gap between remote sensing results and county-level census data to all MODIS pixels within the county.The adjusted remote sensing results based on IAGA not only agree well with the census data in area amount,but also show high agreement with those before IAGA in spatial locations.Such fusion strategy that“Remote sensing is key and census data is secondary”not only can reduce errors from remote sensing,such as training data,data processing and classification method,but also has high immunity to the uncertainty from the census data.