Research On Identification Of Poplar Species In Beijing Area Based On Image Analysis

Plant is the most popular and widely distributed life form on the earth, which influences the crucial ecosystem to human living and provides various indispensable resources for us to survive and develop. Meanwhile, agriculture, which is vital to national economy and development, is closely connected with plant. Therefore, our research on automated plant species identification is of practical significance. The traditional identification relies on the experience of professionals to distinguish plants due to different external morphological characteristics, which is difficult for layman to fulfill in daily life. Moreover, there are approximately 400 thousand species of plants on the earth, which is nearly impossible to be fully identified only counting on memory and experience.In this research, poplar in Beijing area was studied by collecting images of poplar leaves and barks with digital camera and scanner and analyzing shape and textural features with computer, in which image processing and identifying techniques were used to identify poplar species in Beijing. We studied automated plant identification technology based on image analysis, including leave image feature analysis and bark image feature analysis, and made specific improvement in classifiers. Main work and contributions in this research were summarized as follows::(1)A joint identification method based on multi-organs of plant was proposed. Most current researches were based on single plant organ and processing feature selection and extraction. In contrast, we used multiple organs from single plant as the feature sample, extracting and identifying features from both leaves and bark. It improves the sample diversity and raises the average accuracy of identification. The results indicate a significant increase in identification accuracy from multiple organs, leaves and bark, than single organ, and shows a new idea and method on raising the plant identification accuracy.(2)ln the aspect of feature extraction, we designed a specific poplar feature extraction algorithm consisting of the extraction of similarities and dissimilarities of the features of leaves and barks of different popular species, and the imaging classification method. Based on the results from feature selection, according to the demand and feature optimizing results, we chose the leaves shape features (Complexity、Eccentricity、Circularity、Sphericity、Aconvexity、Pconvexity、Aspect ratio、 Rectangularity、Perimeter ratio of diameter、Double lobation、Fractal dimension、Lacunarity、Energy、 Entropy、Contrast、Correlation), leave textual features (Fractal dimension、Lacunarity、Energy、Entropy、 Contrast、Correlation) and the bark textual features (Fractal dimension、Lacunarity、Energy、Entropy、 Contrast、Correlation、Local homogeneity), as the input variables for identifying and separating different poplar species.(3)An improved classifier, named Adaboost.M2-KNN classifier, was put forward based on Adaboost classifier and KNN classifier. Adaboost.M2 classifier was introduced in our study, as the algorithm of original Adaboost classifier requires that the error rate for each weak classifier (or learner) is smaller than 50%. M2 classifier has advantage of higher accuracy by assessing weak classifier through pseudoloss, and punishing wrongly grouped samples and incorrectly identified classifiers, resulting in more training opportunities. Nonetheless, M2 classifier selects weak classifier based on pseudoloss rule, during which the generation of weak classifier is complicated and the training identifying time of algorithm is long. According to the studies mentioned above, KNN classification was selected as the weak classifier for AdaBoost.M2 algorithm, which can significantly reduce the complexity generated by weak classifier, also it will have further improvements in the following aspects: 1.Dramatically reduce the training time and frequency, the average training time of this algorithm is 30.8 sec, which is 5 to 290 sec shorter than the other classifiers; 2.Cut down the average identification time. This algorithm’s average identification time is 0.9 sec and is 0.1 to 19.2 sec shorter than other classifiers; 3.1mprove the average identification rate. This algorithm’s average identification time is 92.93% which is 4% to 11% higher than the other classifiers. In summary, this algorithm has significant improvements in many aspects of classification and identification.