Pouya Ahmadi; Tayebe Managhebi; Hamid Ebadi; Behnam Asghari
Abstract
With the development of remote sensing science, the use of hyperspectral images is becoming more widespread. Classification is one of the most popular topics in hyperspectral remote sensing. In the last two decades, a number of methods have been proposed to address the problem of hyperspectral data classification.In ...
Read More
With the development of remote sensing science, the use of hyperspectral images is becoming more widespread. Classification is one of the most popular topics in hyperspectral remote sensing. In the last two decades, a number of methods have been proposed to address the problem of hyperspectral data classification.In the present study, a structure based on learning capsule networks has been used to classify hyperspectral images, so that the network structure can have the most optimal generation of features by using a convolution layer and a capsule layer, and at the same time Avoid overfitting the on training data. The obtained results show the high quality of production features in the proposed structure. With the development of remote sensing science, the use of hyperspectral images is becoming more widespread. Classification is one of the most popular topics in hyperspectral remote sensing. In the last two decades, a number of methods have been proposed to address the problem of hyperspectral data classification.In the present study, a structure based on learning capsule networks has been used to classify hyperspectral images, so that the network structure can have the most optimal generation of features by using a convolution layer and a capsule layer, and at the same time Avoid overfitting the on training data. The obtained results show the high quality of production features in the proposed structure.In order to improve the classification accuracy, the feature extraction approach through the designed network and the classification by the Extreme Gradient Boosting was compared with the classification method by the global deep network. The proposed capsule approach consists of 3 basic layers: 1) Prime caps, which are capsules of size 8 and 32 with 9 × 9 filters and movement step 2, 2) Digitcaps with 10 16-dimensional capsules, and 3) fully connected layer. The results of examining two approaches for deep networking as well as combining capsule networks with XGBoost reinforcement tree algorithm were compared. Approaches such as SVM, RF-200, LSTM, GRU and GRU-Pretanh were considered to compare the proposed approach based on the configurations mentioned in their research.Up in addition to the study and quality measurement of production vector deep features by the proposed method in different classifiers, the ability of deep global networks in the application of classification should also be examined. The results of examining two approaches for deep network and also combining CapsNet with XGBoost show that by using the proposed combined method, images are classified with 99% accuracy on training data and 97.5% accuracy on test data.Up in addition to the study and quality measurement of production vector deep features by the proposed method in different classifiers, the ability of deep global networks in the application of classification should also be examined.The results of examining two approaches for deep network and also combining CapsNet with XGBoost show that by using the proposed combined method, images are classified with 99% accuracy on training data and 97.5% accuracy on test data.
Mina Hamidi; Hamid Ebadi; abbas kiani
Abstract
By improvement of the spatial resolution of remote sensing images, more accurate information are provided from the image scene such as texture structures. However, extraction of land cover information from these datas has become a challenging process due to the high spectral diversity and the heterogeneity ...
Read More
By improvement of the spatial resolution of remote sensing images, more accurate information are provided from the image scene such as texture structures. However, extraction of land cover information from these datas has become a challenging process due to the high spectral diversity and the heterogeneity of surface materials. Visual interpretation is costly and time consuming and automatic interpretation of images does not necessarily lead to high accuracy. Achieving optimal interpretation accuracy requires the design of automatic algorithms that are capable of dealing with the complexity of the image scene. To overcome this problem, object-based image analysis (OBIA) that is sensitive to the image scene morphology, can be particularly effective in an urban area where the density of man-made structures is high. In object-based classification, pixels of a segment are analyzed in combination with each other. So the dimensions of the problem space are reduced, in compared to the pixel-based method, which leads to increasing the computational speed. Meanwhile, due to the different sizes of image segments, supervised object-based classification faces challenges in creating an optimal training set. In this research, AdaBoost algorithm was selected for the object-based classification, to overcome the problem of feature space imbalance, due to the small number of training samples in comparison with the high dimensions of the feature space (including spectral, spatial and geometric features), two strategies were proposed. In the first approach an active learning mechanism was integrated with AdaBoost to produce optimal training data set (OTD) and in another approach based on the feature-to-feature correlation (redundancy) and the feature-to-class correlation (relevance), the candidate feature subset (CFS) was generated to reduce the size of the feature space. To evaluate the proposed method, the developed algorithm was performed on the standard dataset of Vaihingen in Germany and the results were compared with the pixel-based classification. In order to evaluate the signification of the results, the McNemar statistical test was used. The experimental results showed that the proposed object-based approach improved the overall accuracy by 6% and the kappa coefficient by 7% compared to the pixel-based approach. Also, the computational speed of proposed object-based AdaBoost was significantly increased compared to the pixel-based approach. These results indicate the superiority of the proposed approach both in terms of accuracy and processing speed.
Amir Aghabalaei; Hamid Ebadi; Yasser Maghsoudi
Abstract
Monitoring the earth and its biosphere is an essential task in any scale to achieve a sustainable development. Therefore, forests, as an invaluable natural resource, have an important role to control the climate changes and the carbon cycle. For this reason, biomass and consequently forest height have ...
Read More
Monitoring the earth and its biosphere is an essential task in any scale to achieve a sustainable development. Therefore, forests, as an invaluable natural resource, have an important role to control the climate changes and the carbon cycle. For this reason, biomass and consequently forest height have been known as the key information for monitoring the forest and its underlying surface. Several studies, it has been shown that Synthetic Aperture RADAR (SAR) imaging systems can greatly help to this purpose. In this framework, a novel technique called Polarimetric SAR Interferometry (PolInSAR) is an appropriate and an available tool for forest height estimation, due to its sensitivity to location and vertical distribution of the forest structural components. Based on this, from a view point, the methods employed in this field can be divided into two categories: a) based on Random Volume over Ground (RVoG) inversion model, and b) based on model-based decomposition techniques of PolInSAR data. In this study, in order to improve the forest height estimation, a novel method based on the combination of two mentioned categories has been proposed. The performance and the efficiency of the proposed method were demonstrated by four datasets related to the Pine and the deciduous forests which simulated from the PolSARProSim software in L and P bands.
nima farhadi; Abas Kiani; Hamid Ebadi
Abstract
Object detection is one of the fundamental issues in image interpretation process, especially from remote-sensing imagery. One of the most effective and efficient methods in this field is the use of deep learning algorithm for feature extraction and interpretation. An object is a collection of unique ...
Read More
Object detection is one of the fundamental issues in image interpretation process, especially from remote-sensing imagery. One of the most effective and efficient methods in this field is the use of deep learning algorithm for feature extraction and interpretation. An object is a collection of unique patterns that differ with own adjacent properties. This difference usually occurs in one or more features simultaneously, which can be indicated by the difference in shape, color, and gray values. In this regard, the use of deep learning as an efficient branch of machine learning can be useful in generating high-level concepts through learning in different layers. In this research, a database based on the environmental and geographical conditions from some Iranian airports was created. Additionally, an optimal learner model was developed with a convolutional neural network. For this purpose, in the raw data processing section, besides using the transfer learning method, some vectors were extracted to classify the objects and delivered to an SVM model. The output values were compared with the values obtained from the test image for each object, and they were analyzed in a repeatable process for structural matching. Precision of 98.21% and F1-Measure of 99.1% was achieved, for identification of the target objects
Abbas Kiani; Hamid Ebadi; Hekmat allah Khanlou
Volume 10, Issue 4 , February 2019, , Pages 27-54
Abstract
Land cover classification in remote sensing imagery is one of the most widely used spatial information extraction methods, which can facilitate generating object imagery classes of the ground surface in order to automate and accelerate meeting the basic needs of management, organization, and exploitation ...
Read More
Land cover classification in remote sensing imagery is one of the most widely used spatial information extraction methods, which can facilitate generating object imagery classes of the ground surface in order to automate and accelerate meeting the basic needs of management, organization, and exploitation of the environment. Due to the similar behavior of pixels, remote-sensing image classification using merely the spectral and textural information would lead to inefficiency in the classification. In fact, in classification process, objects are commonly identified using spectral properties of image pixels. If the spatial and conceptual properties are also considered, it causes to a better distinction between image classes and closes the machine process to human interpretation and adds to the system's performance. The present research is mainly focused on the use of interactive segmentation and interpretation processes with respect to the geometry of the image classes. The accuracy of the results have improved by introducing the knowledge-based rules to control and regulate the interactive process, taking into account the geometric properties of target classes. To evaluate the efficiency of the proposed method, the results were evaluated and compared with some of the other methods on IRS satellite images in an urban area. The results showed that geometric and conceptual features as a complementary information source, improve classification results in the urban area with heterogeneous spectral effects. Overall, the proposed hybrid technique improved overall accuracy and Kappa coefficient by 8% and 11.5%, respectively.