Research On Cotton Cultivated Area Identification Algorithm Based On Sentinel-1/2 Satellite Imagery

Cotton is an important cash crop and global strategic material.Accurate and timely acquisition of spatial distribution of cotton cultivated area is of great significance to cotton cultivation management and sustainable development of the cotton industry.Remote sensing has been widely used in crop area identification and classification due to its advantages of wide coverage and dynamic update capability.However,most of the previous studies on cotton identification based on remotely sensed images belong to supervised classification method.The classification accuracy of a single classifier varies with the study area and data source used.There is no classifier that can achieve optimal performance in all cases.Moreover,the supervised classification method is highly dependent on training samples,which limits their application in cotton mapping at a large scale.To address these issues,based on the Sentinel-2 Multispectral Instrument(MSI)and Sentinel-1 Synthetic Aperture Radar(SAR)imagery during the cotton growth period,this study conducted the researches on cotton cultivated area identification algorithm from three aspects,including developing an ensemble learning algorithm by combining multiple classifiers with objective weighting methods to improve the classification accuracy,developing a Cotton Mapping Index to reduce the dependence of samples during the cotton identification process,as well as developing a method based on indices and automatic thresholding method to improve the automatic ability of cotton cultivated area identification.This paper provides a method reference for remote sensing-based cotton identification at a large scale.The main research contents and conclusions are as follows:(1)This study developed an ensemble learning algorithm by combining multiple classifiers with objective weighting methods when the sample data were available.The K-nearest neighbor(KNN),support vector machine(SVM),random forest(RF),back propagation neural network(BPNN)and one-dimensional convolutional neural network(1D CNN)were selected as the base classifiers.The time series of Normalized Difference Vegetation Index derived from Sentinel-2 MSI imagery were used as the input features for five base classifiers.The entropy method and coefficient of variation method were modified,and were combined with the combination weighting method to ensemble the five base classifiers.The results showed that the classification accuracies achieved by the modified weighting methods were higher than those by the original weighting methods.The classification accuracy achieved by combining the modified entropy method and modified coefficient of variation method based on the combination weighting method was slightly higher than those by the single weighting methods.Compared with the five base classifiers,the overall accuracy achieved by the ensemble learning algorithm that based on the normalized combination weighting method with F1-score as input was increased by1.12%～6.45%,0.75%～3.98% and 0.45%～2.70% in three study areas in the Arkansas,Georgia and Texas State of the United States,respectively.Compared with the traditional majority vote,probability fusion and accuracy-weighted methods,both the accuracy and stability of the base classifiers were considered by the proposed ensemble learning algorithm.(2)This study proposed a Cotton Mapping Index(CMI)for cotton mapping by combining spectral and SAR features under the condition that the sample data were unavailable.The spectral and SAR features for cotton and non-cotton crops were analyzed based on Sentinel-2 MSI and Sentinel-1 SAR imagery.The results showed that at the NDVI peak period,the reflectance of red-edge 1 and red-edge 2 for cotton were much higher than those for non-cotton crops,whereas the spectral angle at the red band as well as the absolute values of backscattering coefficients in vertical transmit and vertical receive(VV)polarization for cotton were much lower than those for non-cotton crops.Considering the phenology differences of crops in different regions,the CMI was constructed with the above features,multi-temporal remote sensing imagery and an adaptive window.Then,the cotton cultivated area was identified by combining the CMI and empirical threshold method.The overall accuracy for the four study areas in the Arkansas,Mississippi,Georgia and Texas State of the United States was higher than 81.20%,and the mean relative error in the study areas in Xinjiang of China was 26.69%.Compared with the supervised classification method,the CMI had the advantage of low dependence on samples and no need for training samples in cotton mapping.The CMI,which incorporated optical and SAR features,had a better performance than the indices which using optical features solely.Moreover,the results indicated the potential of the CMI in early season cotton mapping,improving the timeliness of cotton mapping.(3)To solve the relatively high commission error existed in the CMI-derived cotton mapping results with the main error area of peanuts,a Cotton and Peanuts Difference Index(CPDI)was proposed based on spectral angle formed by near infrared band and two short infrared bands of Sentinel-2 MSI imagery.Then,a cotton area identification method based on CMI,CPDI and automatic thresholding method was proposed.The experiments in six study areas in the Arkansas,Mississippi,Georgia,Texas,Alabama and Arizona State of the United States showed that,for the study area with a high percentage of peanuts cultivation,compared with CMI,the commission error was reduced and the overall accuracy was improved when combining the CMI and CPDI for cotton mapping.The spatial threshold method considers the relationship between pixels and adjacent pixels through gray co-occurrence matrix,and performs better than the maximum inter-class variance method,maximum entropy threshold method and triangular threshold method in determining the thresholds of CMI and CPDI.Compared with the empirical threshold method,this method is independent of reference data and is capable of calculating the threshold according to the distribution characteristics of data in different regions,which further improves the automatic level of cotton mapping.The cotton cultivated areas in the study areas in Shandong province of China,India,Pakistan and Uzbekistan were identified using the CMI,CPDI and spatial threshold method.The correlation between the extracted area and statistical area was high.