بهبود دقت طبقه بندی تصاویر ماهواره ای ابرطیفی با استفاده از توسعه یک شبکه عصبی کانولوشن و یادگیری عمیق

اﻣﺮوزه ﺑﺎ ﭘﯿﺸﺮﻓﺖ ﺗﮑﻨﻮﻟﻮژی، ﺳﻨﺠﻨﺪهﻫﺎی ﻣﺘﻌﺪد، ﺗﺼﺎوﯾﺮ اﺑﺮﻃﯿﻔﯽ ﺑﺎ رزوﻟﻮﺷﻦ ﻃﯿﻔﯽ و ﻣﮑﺎﻧﯽ ﺑﺎﻻ ﺗﻬﯿﻪ ﻣﯽﮐﻨﻨﺪ. ﺗﺎ ﮐﻨﻮن روشﻫﺎی ﻣﺘﻌﺪدی ﺑﺮای ﻃﺒﻘﻪﺑﻨﺪی ﺗﺼﺎوﯾﺮ اﺑﺮﻃﯿﻔﯽ اراﺋﻪ ﺷﺪه اﻧﺪ ﮐﻪ ﻫﺮﮐﺪام ﺳﻌﯽ دارﻧﺪ ﺑﺮ ﺗﻌﺪادی از ﭼﺎﻟﺶﻫﺎی ﻣﺤﺎﺳﺒﺎﺗﯽ و ﭘﺮدازﺷﯽ دادهﻫﺎی اﺑﺮﻃﯿﻔﯽ ﻓﺎﺋﻖ آﯾﻨﺪ. ﺑﺎ اﻓﺰاﯾﺶ ﺗﻌﺪاد ﭘﺎراﻣﺘﺮﻫﺎ ﮐﻪ در ﻧﺘﯿﺠﻪ اﻓﺰاﯾﺶ ﺗﻌﺪاد ﻻﯾﻪﻫﺎ در ﺷﺒﮑﻪﻫﺎی ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ رخ ﻣﯽدﻫﺪ، ﮐﺎرآﯾﯽ آﻧﻬﺎ را در ﺣﻞ ﻣﺴﺎﺋﻞ ﭘﯿﭽﯿﺪه ﻣﺎﻧﻨﺪ ﻃﺒﻘﻪﺑﻨﺪی ﺗﺼﺎوﯾﺮ اﺑﺮﻃﯿﻔﯽ ﮐﺎﻫﺶ ﻣﯽدﻫﺪ. در ﺳﺎلﻫﺎی اﺧﯿﺮ، ﻣﻔﻬﻮم ﯾﺎدﮔﯿﺮی ﻋﻤﯿﻖ و ﺑﻪ وﯾﮋه ﺷﺒﮑﻪﻫﺎی ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ، ﺑﻪ دﻟﯿﻞ ﺗﻮﻟﯿﺪ ﺧﻮدﮐﺎر وﯾﮋﮔﯽﻫﺎ و ﮐﺎﻫﺶ ﭘﺎراﻣﺘﺮﻫﺎ ﻧﺴﺒﺖ ﺑﻪ ﺷﺒﮑﻪﻫﺎی ﻋﺼﺒﯽ ﭼﻨﺪﻻﯾﻪ ﭘﺮﺳﭙﺘﺮون ﺑﺎ ﺑﻪ اﺷﺘﺮاک ﮔﺬاﺷﺘﻦ ﭘﺎراﻣﺘﺮﻫﺎ در ﻫﺮ ﻻﯾﻪ، ﺑﺴﯿﺎر ﻣﻮرد ﺗﻮﺟﻪ ﻣﺤﻘﻘﺎن در ﺣﻮزه ﺷﻨﺎﺳﺎﯾﯽ اﻟﮕﻮ ﻗﺮار ﮔﺮﻓﺘﻪ اﺳﺖ. ﻫﺪف ﮐﻠﯽ ﺗﺤﻘﯿﻖ ﭘﯿﺶ رو، ﺗﻮﺳﻌﻪ ﯾﮏ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ ﺑﺮای ﻃﺒﻘﻪ-ﺑﻨﺪی ﺗﺼﺎوﯾﺮ اﺑﺮﻃﯿﻔﯽ اﺳﺖ. ﻧﻮآوری اﺻﻠﯽ اﯾﻦ ﻣﻘﺎﻟﻪ اراﺋﻪ ﭼﺎرﭼﻮﺑﯽ در راﺳﺘﺎی ﺑﮑﺎرﮔﯿﺮی ﯾﺎدﮔﯿﺮی ﻋﻤﯿﻖ ﺷﺎﻣﻞ ﭼﻬﺎر ﻣﺮﺣﻠﻪ ﻣﯽﺑﺎﺷﺪ. ﻣﺮﺣﻠﻪ اول( ﮐﺎﻫﺶ ﺑﻌﺪ ﺑﺎ اﺳﺘﻔﺎده از روش زﯾﺮﻓﻀﺎ، ﻣﺮﺣﻠﻪ دوم( آﻣﺎدهﺳﺎزی ورودیﻫﺎی ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ، ﻣﺮﺣﻠﻪ ﺳﻮم( ﻏﻨﯽﺳﺎزی دادهﻫﺎی آﻣﻮزﺷﯽ، ﻣﺮﺣﻠﻪ ﭼﻬﺎرم( ﻃﺮاﺣﯽ ﻣﻌﻤﺎری ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﮐﺎﻧﻮﻟﻮﺷﻦ. ﭘﯿﺎدهﺳﺎزی ﭼﺎرﭼﻮب ﭘﯿﺸﻨﻬﺎدی ﺑﺮ روی دادهﻫﺎی ﺑﻨﭻ ﻣﺎرک داﻧﺸﮕﺎه ﭘﺎوﯾﺎ، ﻋﻠﯽرﻏﻢ ﺑﮑﺎرﮔﯿﺮی ﺗﻌﺪاد ﻣﺤﺪودی داده آﻣﻮزﺷﯽ، ﻣﻮﺟﺐ ﺣﺼﻮل ﺻﺤﺖ ﻃﺒﻘﻪﺑﻨﺪی 98/3 درﺻﺪ ﺷﺪ.

the accuracy improvement of the hyperspectral satellite image classification by using the development of a convolutional neural network and deep learning

nowadays, with the advancement of technology, numerous sensors provide high spatial and spectral resolution images. so far, several methods have been proposed for hyperspectral image classification, each of which seeks to overcome a number of computational and processing challenges of hyperspectral data. the efficiency of multi-layer perceptron neural networks is greatly reduced due to the increase in the number of parameters along with the increase of the layers, which is essential in complex topics such as hyperspectral image classification. in recent years, the concept of deep learning, especially convolutional neural networks (cnn), has attracted the attention of pattern recognition researchers due to the automatic generation of features and the reduction of parameters compared to the multi-layer perceptron neural networks by sharing the parameters in each layer. the goal of the present study is to develop a convolutional neural networks (cnn in order to classify hyperspectral images. the innovation of this study is to provide a framework to use deep learning. the proposed framework includes four steps. the first step is to reduce dimension by using the sub-space method, the second step is to prepare the cnn inputs, the third step is to augment the teaching data, and the fourth step is to design the cnn architecture. implementation of the proposed framework on the benchmark data of the university of pavia, despite the use of a limited number of educational data, led to the classification accuracy of 98.3%.