Research On Unsupervised Feature Selection And Optimization Algorithm Of Multi-scale Management Zone

It is very important for precision agriculture to establish proper agricultural management zones according to the spatial differences of agricultural input,operation and output.Aiming at the problem of index selection and fragmentation of management zones,this thesis selects part of winter wheat area in China as research regions,applying unsupervised machine learning algorithms to index selection of management zones and optimize the integrity of the management zones on the aspect of index selection.The main contents are as follows:(1)design and establish the multi-scale system,and construct the origin index set of management zones.The design of spatial scale including extend and grain.In the winter wheat areas in China,this thesis established four levels of extend.They are winter wheat area level,province level,city level and county level.Main winter wheat area,Jiangsu,Nantong and Rugao are chosen to be research areas of management zones corresponding to each level.The corresponding grain is established according to the area difference in research areas at all levels and the grain is represented as resolution.The grains of the winter wheat areas level:50km,100km.The grains of province level:10km,25km;The grains of city level:5km,10km;The resolutions of county level:1km,2.5 km.The origin indexes of management zones are identified as meteorological,soil and topographical factors in accordance with the production environment of winter wheat.Meteorological indexes are:cumulative effective sunshine duration(SSD-sum),Growingdegree-days(GDD-sum),daily range of average temperature(DTR-avg),average precipitation(PRE-avg)and cumulative precipitation(PRE-sum).They are calculated and obtained with origin meteorology data based on 3 key growth stages which include jointing stage,flowering stage and maturity stage.Soil indexes are total nitrogen(TN),pH,organic matter(OM),available potassium(AK)and available phosphorus(AP).Topographic indexes are elevation(DEM),slope direction(ASP)and slope(SLO).(2)An unsupervised filtering feature selection algorithm(FSCC)based on correlation clustering algorithm(FSCC)is proposed and applied to the index selection of management zones.In view of the defect that the Variance and Laplacian Score method ignore the correlation between features,FSCC carries out correlation analysis on the origin indexes set to reconstruct the corresponding vector of the indexes,and obtains the final feature index by screening the redundancy indexes of the same cluster through the Affinity Propagation clustering algorithm(AP cluster analysis).At the same time,three kinds of feature selection algorithms are applied to the index selection of management zones in four different scales and their resolutions in this thesis,evaluating advantages and disadvantages of index selection methods.Results show that the FSCC method keeps the index set the diversity of the original features,compared with the Variance method and Laplace Score method,FSCC index set has better performance,in the best performing in Rugao 2.5 km,the evaluation index evaluation indexes fuzzy performance index(FPI),normalized classification entropy(NCE)and the modified separation entropy(MPE)are lower than the other two methods were 52.44%,49.45%and 49.52%.(3)An index set optimization method,CIO,based on spatial consistency and integrity,is proposed for management zones’ fragmentation.In order to improve the practical operability and guide of the management zones,the CIO is proposed to further optimize the results of FSCC indexes set according to the partition fragmentation.The CIO adopts a backward search strategy and coupling Kappa coefficient with FMZ(fragmentation of management zones)to constrain the consistency and integrity before and after partition,further reducing the quantity of element indexes set and removing the index which lead to fragmentation.The CIO optimization experiment is carried out on the corresponding FSCC index set in the research area at each scale and resolution and the experimental results are analyzed.Compared to FSCC,In Rugao 1 km,Rugao 2.5km,Nantong 5 km and Nantong 10 km,CIO has a substantial improvement on the integrity of management zones that total quantities of patches are reduced by 98.8%,91.5%,90%,83.9%respectively.Except Nantong 10 km,CIO is 0.078,0.061 and 0.082 lower than FSCC index set in FPI,NCE and MPE respectively,which relatively improved the partition effect of FSCC.The research and its result in this thesis have certain reference value for the index selection of multi-scale management zones,and provide reference and convenience for the formulation of multi-scale management zoning and management zone scheme under the current agricultural technology extension system and the reform of the form of agricultural technology personnel service.