بهبود تشخیص صحنه در سنجش از راه‌دور با استفاده از یادگیری عمیق و انتخاب‌گر ویژگی

تصاویر سنجش‌‌ازدور به‌عنوان منبع‌داده با ارزش می‌تواند در اندازه‌گیری و مشاهده ساختارهای دقیق در سطح زمین کمک کند. هدف این پژوهش ارائه راهکار تشخیص صحنه در سنجش از راه‌‌دور با به‌کارگیری روش‌های یادگیری عمیق است. برای رفع محدودیت‌های روش‌های پیشین، یک رویکرد ترکیبی استخراج ویژگی مطرح شده‌است که در آن دو نوع ویژگی عمیق محلی و سراسری و یک ‌نوع محلی دستی با یکدیگر ترکیب شده اند. برای استخراج ویژگی‌ها یک شبکه‌ کانولوشن پیش‌آموزش‌دیده با بیست لایه تمام‌متصل پیشنهاد شده‌است؛ همچنین یک مرحله انتخاب ویژگی با استفاده از دو دسته الگوریتم‌های پالایه و بسته بند در این ‌مدل تعبیه شده‌است؛ در نهایت با استفاده از الگوریتم‌های طبقه‌بندی مختلف تشخیص صحنه انجام می‌شود. ارزیابی راهکار پیشنهادی برای مجموعه‌داده های ucm، aid ،rsscn7 و nwpu-resisc45 به‌ترتیب دقت 99.27% ، 97.91% ، 99.09% و 93.09% را کسب کرده‌است.

improving scene recognition in remote sensing using deep learning and feature selector

remote sensing images as a valuable data source in earth observation can help in measuring and observing detailed structures on the earth’s surface. scene detection in remote sensing images has many applications in various fields such as urban planning, natural hazard detection, environmental monitoring, vegetation mapping, and geographic object detection. one of the key problems in the interpretation of remote sensing images is the scene classification of remote sensing images. feature extraction is very important in scene detection and classification. convolutional neural networks are one of the deep learning methods that have significantly increased the performance of tasks such as object recognition and scene classification, but their performance is highly dependent on the number of labeled images available, which are not available enough especially in the field of remote sensing. recently, transfer learning, especially for the fine-tuning of pre-trained convolutional neural networks, has attracted more attention from researchers as a practical strategy for scene classification in remote sensing. however, the lack of use of local features and global deep model that is trained on the target data set is one of the limitations of current methods. also, if these networks are not deep enough and the images do not pass through multiple filters, they cannot extract more semantic information, and the extracted features do not have high discrimination power, and as a result, scene recognition is not performed well. on the other hand, the features extracted through local features are very large, and not using feature selector methods reduces the accuracy of the model. in this research, to solve the mentioned limitations, a hybrid approach of feature extraction has been proposed in which three types of features including two types of deep local and global features and one type of manual local feature are combined with each other. to extract deep features, pre-trained convolutional networks have been used. the pre-trained networks used are: resnet, inceptionnet, googlenet and efficientnet_b0. in order to extract as much information as possible from the images, a convolutional network with 20 fully connected layers is proposed. also, a combined feature selection stage consisting of two categories of filtering and packing algorithms is included in this model. finally, scene detection is performed using several different classification algorithms. the different structure of pre-trained convolutional networks and their appropriate depth can be effective in improving the extraction of deep features. in addition, the combination of three categories of different features can provide a more comprehensive knowledge of images. the evaluation of the proposed solution on the ucm, aid, rsscn7 and nwpu-resisc45 datasets has obtained the accuracy of 99.27%, 97.91%, 99.09% and 93.09% respectively in identifying and classifying images. as a result, this solution has shown a better performance compared to the models that used the manual extraction of features, as well as the methods that use normal convolutional models.