اﻧﺘﺨﺎب ﻃﻮل ﻣﻮج ﻫﺎي ﻃﯿﻔﯽ ﺣﺴﺎس از ﻣﯿﺎن داده ﻫﺎي ﻓﺮاﻃﯿﻔﯽ ﺟﻬﺖ ﺗﺸﺨﯿﺺ ﺑﺮگ ﻫﺎي درﺧﺖ ﭘﺴﺘﻪ آﻟﻮده ﺑﻪ ﭘﺴﯿﻞ

Spectral Feature Selection from the Hyperspectral Dataset to Identify Pistachio Leaves Infected by Psylla

در ﺳﺎل ﻫﺎي اﺧﯿﺮ ﺗﻮﻟﯿﺪ ﭘﺴﺘﻪ ﺗﻮﺳﻂ آﻓﺘﯽ ﺑﻪ ﻧﺎم ﭘﺴﯿﻞ ﺗﻬﺪﯾﺪ ﺷﺪه اﺳﺖ. ﻫﺪف از اﻧﺠﺎم اﯾﻦ ﺗﺤﻘﯿﻖ اﻧﺘﺨﺎب ﺑﺎﻧﺪﻫﺎي ﻃﯿﻔﯽ ﻣﻨﺎﺳﺐ ﺟﻬـﺖ ﺗﺸـﺨﯿﺺ ﺑﺮگ ﻫﺎي آﻟﻮده ﺑﻪ آﻓﺖ ﭘﺴﯿﻞ ﻣﯽ ﺑﺎﺷﺪ. ﺑﺮاي اﯾﻦ ﻣﻨﻈﻮر از 160 ﺑﺮگ ﺳﺎﻟﻢ و 160 ﺑﺮگ ﺑﯿﻤﺎر در 64 ﺑﺎﻧﺪ ﻃﯿﻔﯽ در ﻧﺎﺣﯿﻪ ﺑـﯿﻦ 400 ﺗـﺎ 1100 ﻧـﺎﻧﻮﻣﺘﺮ ﺑـﺎ ﻗﺪرت ﺗﻔﮑﯿﮏ ﻃﯿﻔﯽ ﺣﺪوداً 10 ﻧﺎﻧﻮﻣﺘﺮ ﻃﯿﻒ ﻧﮕﺎري اﻧﺠﺎم ﺷﺪ. از روش ﺟﻨﮕﻞ ﻫﺎي ﺗﺼﺎدﻓﯽ ﺑﺮاي ﻣﺮﺗﺐ ﺳﺎزي ﻃﻮل ﻣﻮج ﻫﺎ ﺑﺮ اﺳﺎس ﻗـﺪرت ﺗﻔﮑﯿـﮏ ﺑـﯿﻦ ﺑﺮگ ﻫﺎي ﺳﺎﻟﻢ و ﺑﯿﻤﺎر اﺳﺘﻔﺎده ﺷﺪ. ﭘﺲ از ﺷﻨﺎﺳﺎﯾﯽ و ﻣﺮﺗﺐ ﺳﺎزي ﻃﻮل ﻣﻮج ﻫﺎ ﺑﺮ اﺳﺎس ﻣﯿﺰان اﻫﻤﯿﺖ، ﻃﻮل ﻣﻮج ﻫﺎي داراي ﺑﯿﺸﺘﺮﯾﻦ اﻫﻤﯿﺖ ﺗﻮﺳﻂ ﯾـﮏ اﻟﮕﻮرﯾﺘﻢ ﺧﻮﺷﻪ ﺑﻨﺪي در ﺷﺶ ﺧﻮﺷﻪ ﺗﻘﺴﯿﻢ ﺑﻨﺪي ﺷﺪﻧﺪ ﺑﻪ ﻃﻮري ﮐﻪ ﻣﯿﺎﻧﮕﯿﻦ ﻃﻮل ﻣﻮج ﻫﺎي ﻫﺮ ﺧﻮﺷﻪ ﻫﺎ ﺣﺪاﻗﻞ 50 ﻧﺎﻧﻮﻣﺘﺮ از ﯾﮑﺪﯾﮕﺮ ﻓﺎﺻﻠﻪ داﺷﺘﻪ ﺑﺎﺷـﻨﺪ. ﻣﯿﺎﻧﮕﯿﻦ ﻃﻮل ﻣﻮج ﻫﺎي ﻗﺮار ﮔﺮﻓﺘﻪ در ﺷﺶ ﺧﻮﺷﻪ ﺑﻪ ﺗﺮﺗﯿﺐ اﻫﻤﯿﺖ ﻋﺒﺎرت ﺑﻮد از: 468 ،598 ،858 ،791 ،710، و 1023 ﻧـﺎﻧﻮﻣﺘﺮ. ﺑـﺮاي ﺗﺒـﺪﯾﻞ داده ﻫـﺎي ﻓﺮاﻃﯿﻔﯽ ﺑﻪ ﭼﻨﺪﻃﯿﻔﯽ، از ﺑﺎزﺗﺎﺑﺶ ﻫﺎي ﺑﻪ دﺳﺖ آﻣﺪه در ﻓﺎﺻﻠﻪ 15 ﻧﺎﻧﻮﻣﺘﺮي اﯾﻦ ﻣﺮاﮐﺰ ﻣﯿﺎﻧﮕﯿﻦ ﮔﯿﺮي اﻧﺠﺎم ﮔﺮﻓﺖ و داده ﻫﺎي ﺑﺎزﺗﺎﺑﺶ ﺑﻪ دﺳﺖ آﻣـﺪه از دﯾﮕﺮ ﻃﯿﻒ ﻫﺎ ﺣﺬف ﺷﺪ. اﻟﮕﻮرﯾﺘﻢ ﻣﺎﺷﯿﻦ ﺑﺮدار ﭘﺸﺘﯿﺒﺎن ﺑﺮاي ﻃﺒﻘﻪ ﺑﻨﺪي ﺑﺮگ ﻫﺎي ﺳﺎﻟﻢ و ﺑﯿﻤﺎر ﺑﺎ اﺳﺘﻔﺎده از داده ﻫﺎي ﻓﺮاﻃﯿﻔﯽ و ﭼﻨﺪ ﻃﯿﻔﯽ ﺑﻪ دﺳﺖ آﻣﺪه در اﯾﻦ ﭘﮋوﻫﺶ ﺑﻪ ﮐﺎر ﮔﺮﻓﺘﻪ ﺷﺪ. دﻗﺖ ﻃﺒﻘﻪ ﺑﻨﺪي ﺑﺎ اﺳﺘﻔﺎده از ﺗﻤﺎم 64 ﻃﻮل ﻣﻮج )داده ﻫﺎي ﻓﺮاﻃﯿﻔﯽ( 90/91 درﺻﺪ و ﺑﺎ اﺳﺘﻔﺎده از 6 ﻃﻮل ﻣﻮج )داده ﻫﺎي ﭼﻨﺪﻃﯿﻔﯽ( 88/69 درﺻﺪ ﺑﻮد. اﺧﺘﻼف ﺑﺴﯿﺎر ﮐﻢ )ﺣﺪود 2 درﺻﺪ( در ﻣﯿﺰان دﻗﺖ ﻃﺒﻘﻪ ﺑﻨﺪي ﻧﺸﺎن دﻫﻨﺪه ﺻﺤﺖ ﺷﯿﻮه اراﺋـﻪ ﺷـﺪه در اﯾـﻦ ﭘـﮋوﻫﺶ ﺑـﺮاي ﮐﺎﻫﺶ اﺑﻌﺎد داده ﻫﺎي ﻓﺮاﻃﯿﻔﯽ ﻣﯽ ﺑﺎﺷﺪ. ﻋﻼوه ﺑﺮ ﮐﺎﻫﺶ اﺑﻌﺎد داده، ﺗﻌﯿﯿﻦ ﺑﺎﻧﺪﻫﺎي ﻃﯿﻔﯽ ﻣﻨﺎﺳﺐ از ﻣﯿﺎن داده ﻫﺎي ﻓﺮاﻃﯿﻔـﯽ ﮔـﺎﻣﯽ ﻣـﻮﺛﺮ در ﻃﺮاﺣـﯽ و ﺳﺎﺧﺖ ﺣﺴﮕﺮي ﭼﻨﺪﻃﯿﻔﯽ ﺟﻬﺖ ﺗﺸﺨﯿﺺ ﺑﯿﻤﺎري ﮔﯿﺎﻫﺎن ﻣﯽ ﺑﺎﺷﺪ.

Pistachio production has been adversely affected by Psylla, which is a devastating insect. The primary goal of this study was to select sensitive spectral bands to distinguish pistachio leaves infected by Psylla from healthy leaves. Diagnosis of psylla disease before the onset of visual cues is crucial for making decisions about topical garden management. Since it is not possible to diagnose psylla disease even after the onset of symptoms with the help of color images by drones, hyperspectral and multispectral sensors are needed. The main purpose of this study was to extract spectral bands suitable for distinguishing healthy leaves from psylla leaves. For this purpose, in this paper, a new method for selecting sensitive spectral properties from hyperspectral data with the high spectral resolution is presented. The intelligent selection of sensitive bands is a convenient way to build multispectral sensors for a specific application (in this article, the diagnosis of psylla leaves). Knowledge of disease-sensitive wavelengths can also help researchers analyze multispectral and hyperspectral aerial images captured by satellites or drones. Materials and Methods A total number of 160 healthy and diseased leaves were scanned in 64 spectral bands between 400-1100 nm with 10 nm spectral resolution. A random forest algorithm was used to identify the importance of features in classifying the dataset into diseased and healthy leaves. After computing the importance of the features, a clustering algorithm was developed to cluster the most important features into six clusters such that the center of clusters was 50 nm apart. To transfer the hyperspectral dataset into a multispectral dataset, the reflectance was averaged in spectral bands within ±15 nm of each cluster center and achieved six broad multispectral bands. Afterwards a support vector machine algorithm was utilized to classify the diseased and healthy leaves using both hyperspectral and multispectral datasets. Results and Discussion The center of clusters were 468 nm, 598 nm, 710 nm, 791 nm, 858 nm, and 1023 nm, which were calculated by taking the average of all the members assigned to the individual clusters. These are the most informative spectral bands to distinguish the pistachio leaves infected by Psylla from the healthy leaves. The F1-score was 90.91 when the hyperspectral dataset (all bands) was used, while the F1-score was 88.69 for the multispectral dataset. The subtle difference between the F1-scores indicates that the proposed pipeline in this study was able to select appropriately the sensitive bands while retaining all relevant information. Conclusion The importance of spectral bands in the visible and near-infrared region (between 400 and 1100 nm) was obtained to identify pistachio tree leaves infected with psylla disease. Based on the importance of spectral properties and using a clustering algorithm, six wavelengths were obtained as the best wavelengths for classifying healthy and diseased pistachio leaves. Then, by averaging the wavelengths at a distance of 15 nm from these six centers, the hyperspectral data (64 bands) became multispectral (6 bands). Since the correlation between the wavelengths in the near-infrared region was very high (more than 95%), out of the three selected wavelengths in the near-infrared region (710, 791, and 1023), only the 710-nm wavelength, which was closer to the visible region, was selected. The results of classification of infected and diseased leaves using hyperspectral and multispectral data showed that the degree of classification accuracy decreases by about 2% and if only 4 bands are used, the degree of accuracy decreases by about 3%. The results of this study revealed that the proposed framework could be used for selecting the most informative spectral bands and accordingly develop custom-designed multispectral sensors for disease detection in pistachio. In addition, we could reduce the dimensionality of the hyperspectral datasets and avoid the issues related to the curse of dimensionalitylity.