Research On Methods For Agricultural Image Recognition And Spectral Detection With Machine Learning

The methods of agricultural data analysis based on machine learning provide the basis for the development of the intelligent agricultural system.Image and spectrum are important perception information of intelligent agricultural machinery.Intelligent analysis and decision making based on acquired image data and spectral data is the key to the intellectualization of agricultural equipment.For these two different kinds of agricultural data,machine learning is an effective way to complete the analysis task.In view of the shortcomings of traditional machine learning,methods of image recognition and spectral detection based on deep learning wree proposed in this paper.The research object is image recognition of crop/weed in field and spectral detection of soil moisture content.Based on the advantages(autonomous mining and learning complex data)of deep learning,the image recognition and spectral analysis models with higher performance were build and optimized,and the basic research problems of deep learning applied in agricultural image and spectral data processing was explored.It provides a new method for agricultural image recognition and spectral detection based on deep learning,and has a certain reference value in both theoretical and practical aspects.The main research work and conclusions are as follows:(1)It has been found that maize seedlings are generally higher than the weeds during the same period,this fact could provide more accurate recognition evidence for the SVM recognition model.The collected images were pretreated with methods of Excess-green feature,Improved Otsu,Area filtering,Canny edge detection.16 morphological features and 2 texture features in monocular image based on image pretreatment were obtained.The height feature extraction method of plant was proposed in this research.Based on binocular images,the height feature of plant can be extracted by this method quickly.The error between actual height and calculation was less than ± 12mm.4)The weeding period could be divided into three stages and SVM recognition model that fusion height feature and image features for each stage was built.Based on the max-min ant system algorithm,the morphological features of each model were optimally selected,and the morphological features were reduced from 16 to 6,feature data was reduced by 62.5%.By using three kinds of algorithms:genetic algorithm,k-fold cross validation,and particle swarm optimization,the two core parameters of each SVM recognition model were optimized,and the optimal parameters were selected by comparing effects df three algorithms.The optimal parameters c and g contained by this way can effectively improve the recognition ability of SVM model,and avoid the over learning and under learning state of SVM model.It is necessary to segment the overlapping images of maize and weed.Watershed algorithm based on distance transform was used to solve this problem.The results of this research showed that the recognition rate of the SVM model based on fusion height feature was 96.67%,100%,98.33%,and the average recognition rate was 98.33%.The data showed that the SVM recognition model based on fusion height feature is better than the model without fusion height feature,and the average recognition rate is improved by 5%.Therefore,weed recognition method based on fusion height feature and SVM model can effectively improve the recognition rate and achieve high accuracy.(2)In order to solve the main problems in the current research,we explored the way to improve the recognition accuracy,stability and real-time performance,and a recognition method of crop and weed based on multiscale CNN model was proposed.The main research contents of this paper were included as follows.Excellent internal features of image are hierarchical.In this research,the multiscale hierarchical feature is a scene level feature with invariance and consistency in scale space.Multiscale CNN was built to extract multiscale hierarchical feature.Multiscale CNN contains 3 multiple copies of a single CNN that were applied to different scales of a Gaussian pyramid of the input image,each CNN model consists of 3 stages.The first 2 stages contain a bank of convolution kernels,a point to point nonlinear mapping activation function,and a spatial pooling module.The last stage only contains a bank of kernels.With completely training,this CNN model could automatically extract hierarchical feature representations from the input image,thereby decreasing the need for hand-engineered features.A series of feature maps for multiscale regions in the image was produced,by the multiscale CNN,and multiscale hierarchical feature is learned to allow recognition the class of all pixels in the image.The average pixel target recognition rate is 93.41%.We used the target region recognition strategy of multiscale CNN combined with the superpixels segmentation in this study.Firstly,an over segmentation of the original image was produced through the SLIC superpixels segmentation method.At the same time,each pixel location of the image was classified,based on the multiscale hierarchical features.These predictions of pixels in each superpixel were aggregated to produce the class prediction of superpixel,through computing the average class distribution within the superpixel.Adjacent superpixels with the same class were merged to obtain the final target class prediction and image segmentation.The accurate image segmentation was produced while recognizing the target region in the image by this recognition strategy,effectively avoid the problems caused by targets overlapping,and more stable and accurate recognition results were achieved.The results of experiments showed that the average target recognition rate of method in this paper was 98.92%,the standard deviation was 0.55%,and the average time to recognize a single image was 1.68 s,thus the stable and accurate results of recognition were achieved.Compared with the traditional machine learning model,the accuracy,stability and time consumption of recognition results by method in this paper were improved in different degrees.The real-time performance of method in this paper can be further improved by GPU hardware acceleration,and the average time to recognize a single image was only 0.72 s.Therefore,recognition method based on multiscale CNN combined with superpixels segmentation can effectively achieve accurate,stable and efficient recognition of crop and weed.(3)In this research,the CNN was used to predict the soil moisture content by near infrared spectroscopy.An efficient modeling method of CNN for spectral regression was proposed.In this paper,firstly,the simple pretreatment were used to treat the spectral reflectance data of soil samples under different moisture contents.Principal component analysis was used to reduce the amount of spectral data and the correlation of the features.The processed spectral data was transformed into 2-dimensional spectral information matrixes to meet the special learning structure of CNN.Secondly,the CNN structure was used to build the regression model for the prediction of soil moisture content.The first four stages of this CNN model are composed of two types of layers:convolutional layers and pooling layers.Inner features of the input spectral data were obtained by composing convolutional layers and pooling layers,each transforms the representation at one level into a representation at a higher,slightly more abstract level.With the composition of enough such transformations,very effective inner features of spectral data can be extracted.The model structure and parameters were optimized by carrying out experiments.Finally,the CNN model with improved regression structure of soil spectral data was built for the prediction of soil moisture content.The CNN model was compared with the BP,PLSR and LSSVM models,and these three traditional models were commonly used in the prediction of soil moisture content.The results shown that when the number of training samples reached to some degree,the prediction accuracy and regression fitting degree of the CNN model were higher than those of the traditional models.The prediction accuracy of the CNN model greatly increased with the number of training samples growing,so did the regression fitting degree of the CNN model.In the end,the performance of the CNN model is significantly better than the traditional models.Therefore,the CNN method could be used to effectively predict the soil moisture content by the near infrared spectral data,and better results are obtained when more training samples are involved in modeling.