Study On Per-field Crop Classification Method Using Remote Sensing Data

Precise spatial distribution of crops is an important basis for further growth monitoring and yield estimation.Remote sensing technology with high speed,wide range and low cost is widely used in crop classification which is the key to extracting crops spatial distribution.The crop classification using remote sensing data can be grouped as pixel-based,object-based and field-based according to its mapping unit.Different approaches has their own advantages,but there are also some problems.It is difficult to overcome the “pepper and salt” phenomenon and the mixed pixel problem for pixel-based classification.Object-based classification can avoid the above problems but it is hard to segment images to objects accurately.As a special object,the crop field has certain advantages as a classification unit,and the nationally released field data set also brings new opportunities for per-field crop classification with remote sensing.Per-field classification can produce more accurate crop distribution maps by mapping crops based on field boundary.However,further studies are needed on the classification features,classification strategies and remote sensing data selection in per-field crop classification.In view of the problems existing in per-field crop classification using remote sensing,this paper took six crop types in the study area as the research object,and used multi-temporal Landsat8 data as the main data source to carry out research on the method of per-field crop classification using remote sensing data.In this study,the relationship between features and crop classification accuracy was analyzed on the basis of feature importance.At the same time,the strategies of per-field classification were developed to overcome its limitation of training samples and problem of mixed pixels.The similarities and differences between per-field classification and object-based classification were also discussed from two aspects: classification accuracy and acreage accuracy.Finally,per-field crop classifications using images with different spatial resolution were implemented to guide data selection in further crop identification.The main studies and conclusions are listed as follows:(1)Feature selection is an important step in classification with remote sensing data.Firstly,the impact of features on per-field crop classification was analyzed using random forest based on the importance scores of 304 features,which came from 7 phases of Landsat8 images.The results indicate that May 10,September 15 and July 29 are the best phases for the classification,and their corresponding spectral bands and vegetation indices are of higher importance.Increasing feature quantity in the order of importance will improve crop classification accuracy.When the number increases to 10,the classification accuracy reaches a steady state.We also found that too much features will not cause over-fitting with random forest.Finally,in the study area,using early images(May 10 and June 27)for crop identification can achieve a rather high overall accuracy(OA)of 87.75%.(2)Taking into consideration that per-field crop classification may have small account of training data and several mixed pixels on field boundary,the strategies of per-field classification was improved and the results of different strategies were compared and analyzed.Taking field as a whole for feature extraction and using it as input for the classifier,strategy A yielded 88.24% OA on testing dataset.However,it had a poor performance in crop types with fewer training fields.With strategy A1,training fields were more presentative by excluding mixed pixels on field boundary and its OA increased to 90.69%.Strategy B and B1 assigned a field to a particular class according to a per-pixel classification.Therefore,the amount of training data increased by transforming fields to pixels,which brought a higher OA of 93.63%.Even though strategy B1 excluded untrustworthy pixels when assigning final class for the field,it did not differ from strategy B due to its majority voting rule and large field.(3)Object-based crop classification was carried out in our study area to analyze similarities and differences between object-based and per-field classification.This paper compared the two methods from the aspects of classification accuracy and acreage accuracy.In terms of classification accuracy,object-based crop classification can reach 88.24% OA,the same as per-field classification using strategy A.In terms of acreage accuracy,the methods showed a similar pattern that the acreage accuracy were consistent with its classification accuracy.However,per-field crop classification yielded a higher acreage accuracy owing to the auxiliary field data,which contains more accurate crop planting acreage information.(4)This paper studied the relationship between spatial resolution and the accuracy of per-field crop classification.The results show that there are a downtrend in OA as a whole with the decrease of spatial resolution.The per-field crop classification get the highest OA of 93.63% when spatial resolution is 15 m and can reach a rather high OA of 84.8% with 60 m spatial resolution data.As a major crop type,the identification of cotton are less affected by spatial resolution.The maximum difference of F1 measure,which is the weighted harmonic mean of user accuracy and producer accuracy,is 14.35%.The lowest F1 measure can achieve 82.54% when image spatial resolution is not lower than 150 m.