Study On Extraction And Classification Of Crop Characteristic Band Based On Spectral Imaging Technology

Monitoring growth status, diseases and pests for crops is one of particularlyimportant steps in the development of agricultural production, which has an impacton yields and qualities of crops. The methods of monitoring crops based on traditionalhuman visual system and RGB digital color equipment are often low efficiency andhigh misjudgment error rate. However, spectral imaging technology integration ofimages and spectrums has solved "images and spectrums separately","metamerism"and other problems in the traditional science fields. In view of this, this paper willmake research on feature band extraction and classification for crops using spectralimaging technology.Firstly, according to the principle of spectral imaging technology, throughanalyzing the advantages and disadvantages of combination of core imaging devicesfor monochrome industrial CMOS camera and liquid crystal tunable filters (LCTF),considering the portability and stability of imaging system, and using effectiveaperture for LCTF uttermost, the experiment builds multi-spectral imaging system forcrops by locating LCTF between lens and CMOS camera, and the paper introducesthe working principle, system calibration, multi-spectral image acquisition, dataprocess and expression for the multi-imaging system.Secondly, the experiment acquires multi-spectral images of four kinds of cropsamples using multi-spectral imaging system from band400nm to720nmwavelength range with an interval of5nm, and spectral feature curves of four kindsof samples are then derived. Considering high redundancy, strong correlation amongeach band, taking up much storage space and enhancing the algorithms difficulty fordata process, so it is necessary to select effective feature bands from a large numberof multi-spectral imaging data for crops. Combined with the selection criterion offeature bands, it is respectively calculated that the result of feature bands for healthyphaseolus vulgaris leaves through the principle of band index and image brightnessinformation. By analyzing characteristics for plants, correlation among each band andthe amount of information, it can be concluded that the result of feature bands for healthy phaseolus vulgaris leaves is ideal through band index. Therefore, theexperiment selects feature bands for other three kinds of samples by the principle ofband index.Finally, according to the method of nearest neighbor, the experiment classifiesfour kinds of samples from the view of whole bands and feature bands respectivelyusing distance method, spectral angle match and correlation coefficient. Experimentalresults show that the accuracy of three kinds of classification methods can achieveideal classification result, which classification accuracy is greater than97.00%forwhole bands. Instead, as to feature bands for crop samples, three kinds ofclassification methods are not with a high precision accuracy, but it can achieve idealclassification accuracy by combination of classification methods for four kinds ofsamples; as to damaged rice leaves by planthoppers, it can also improve classificationaccuracy via increasing the number of feature bands. With regard to the correctclassification for experimental samples, it can be established multi-spectral databaseof feature bands for corresponding crops.Experimental results show that multi-spectral imaging technology can be appliedto monitor growth status as well as diagnose diseases and pests for crops, which is arapid and non-destructive monitoring technology, the experiment applies effectivefeature bands to provide a fast identification way and optimize multispectral imagedata for crops, so it is substantially expected to be applied to identify other cropsquickly, classify precision agriculture, and detect objects by hyperspectral remotesensing.