Temporal And Spatial Characteristics Of Crop Identification Features In Jiangsu Province

Crop area has become a key issue that affects China's grain production.With the macro,rapid monitoring capabilities at large scale,remote sensing technology,can obtain objective and accurate agricultural information in a short time,and has been wildly applied to crop identification and area estimate.At present,three problems must be solved in this aspect to achieve satisfying results.Firstly,most researches involving crop mapping using remotely sensed images had just limited to spectrum reflectance.Consequently,comprehensive features about different types of crops can't be taken full advantage of.Secondly,since the crop identification features have temporal and spatial characteristics,thus they perform variously to the crop classification results with the changes of requisite-date or spatial resolution of remotely sensed images.However,few researches focus on the temporal or spatial effects of these crop identification features before,resulting in the blind selection of feature in terms of date-phase and spatial resolution.Thirdly,crop planting system and land fragmentation vary greatly in different regions of China,even many regions out China.For example,in the south of China,arable land is the small and broken,and the planting structure is complex and diverse,which make crop identification much more difficult than that in north China.Facing too much mixed spectrum pixels to be identified,medium or low resolution remotely sensed data has trouble in discriminating the main crops.Similarly,the phenomenon that the same objects have different spectrum is more serious in high-resolution images,which lowers the classification accuracies.In order to solve these problems,this study focused on the following two aspects:i)based on features synthesis and optimally selected features of multi-temporal remotely sensed data,the effects of accuracy improvements on crop identification was analysis,and the best date-phase and the optimal features subset were determined to be existed for certain crops identification;ii)based on features synthesis and optimally selected features of multi-scale remotely sensed data,the laws were studied between crop identification accuracies and different spatial resolutions,and the appropriate spatial resolution and the optimal features subsets corresponding to each spatial resolution were determined for certain crops identification.This research could be valuable to provide more reliable information to national agriculture and food departments and support their macro decisions.According to the research goals,two experiments were designed.i)This experiment focused on features synthesis and optimally selected features of multi-temporal remotely sensed data.In this experiment,Sihong County was chosen as experiment area,and corn and rice were objects.Firstly,multi-source features were used to construct crop identification features sets,including spectrum,vegetation indexes(VIs),band differences of different time(BDs)and texture features.Next,univariate feature selection(UFE)algorithm was used to select features.Finally,random forest(RF)classifier was applied to discriminate the two main crops based on the feature combinations.ii)The experiment focused on multi-source features and optimally selected features based on multi-scale remotely sensed data.Gaochun County was selected as the experiment area,and the wheat and canola were targets.Firstly,spectrum,VIs and textures were used to construct crop identification features sets.Then,feature recursive elimination(FRE)was used to select features.Finally,the classifier RF and object-oriented classification method were applied to discriminate the main crops based on the selected features.The conclusions of this paper are shown as follows:For the experiment in Sihong County,the importance of different features to crop identification varied greatly and there existed a subset of optimal features,which could improve crops classification accuracies.More specifically,there are four important points found.i)Different features perform differently.The spectrum,VIs and BDs features showed high importance,while texture was less useful.So the new feature,BD showed its feasibility and effectiveness.ii)Similarly,different phases also had different importance.And October 12 th was the appropriate phase when identifying the autumn crops in Sihong County,followed by August.iii)Multi-source features improved crop identification accuracies and each kind of features would contribute to the crop classification.Besides,there was a subset of appropriate features to identify corn and rice indeed.For the two crops in Sihong County,the overall accuracy and Kappa coefficient of this subset achieved 0.9707 and 0.96 respectively.iv)UFE performed well in identification feature selection of crops.For the other experiment in Gaochun County,features' importance to crop identification was different at different spatial resolutions,resulting in the difference in the optimal features subsets.And there was a critical point for spatial resolution selection in using the object-oriented classification or the pixel-based classification.There are six crucial points existing.i)The classification accuracies were not proportional to the number of features.ii)There was a subset of optimal features at any spatial scales,and the feature types and number varied as the spatial resolutions changed.iii)There also existed a subset of shared features suitable for all spatial scales,which was useful to construct the feature sets.iv)With the decrease of the spatial resolution,the classification accuracy went up at first and went down then,suggesting the existence of an ideal spatial resolution.And 10 m was the appropriate spatial resolution to the identification of winter wheat and canola in Jiangsu Province.v)There was a critical point for spatial resolution in using the object-oriented classification or the pixel-based classification.When a spatial resolution was higher than it,the object-oriented classification performed better.And 6m was the critical point in Jiangsu Province.vi)The new VI feature,AreaVI,was of great significance to winter wheat and canola discrimination.