Algorithm and Image Processing
EXTRACTING ROCK INFORMATION BASED ON INTEGRATED CONVOLUTIONAL NEURAL NETWORK
AND IMAGE PROCESSING TECHNOLOGY
Jiangying Wang1, Yun Zhou2 and Miaofen Huang3*
1 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China, Email: 1361748619@qq.com
2 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China, Email: 1282854954@qq.com
3 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China, * Corresponding author: hmf808@163.com
ABSTRACT
The lithology identification and the extraction of oil-bearing information on rock surface are the most critical works in exploring minerals, oil and gas resources. An automatic classification method of integrated convolutional neural network was proposed to identify the lithology of rocks.
Comprehensive image-processing techniques are used to extract the oil- bearing information on the rock surface. Firstly, five convolutional neural network models, including AlexNet, VGG16, Inception_V3, Xception and ResNet50, were trained with the enhanced rock sample set to obtain five different classification models. An integrated model was further established by using the majority voting method; Secondly, the rock area in the image, which was taken in a dark box illuminated by fluorescent light, was extracted by the comprehensive image-processing techniques, such as linear stretching, Canny edge detection, and image expansion, and verification was made with the area of interest extracted by ENVI; Finally, the ratio of oil-bearing area to the area of the rock was calculated by the image masking and HSV transformation processes. The classification results show that the verification accuracy of a single model (Inception_V3) is 98.87%, while the verification accuracy of the integrated model (Xception, VGG16, Inception_V3) is 99.29%, which is better than single model; In addition, the deviation of the rock profile obtained by the comprehensive image-processing techniques fall within 1% compared with the area of interest drawn by ENVI; The percentage of the pixel number of oil-bearing derived from the results processed by the image mask and HSV transformation processes to the total pixel number of the rock area is 2.85%. Compared with the traditional observation procedure, the workflow used in the paper can save time greatly to extract the oil-bearing information.
Keywords: Lithology Identification; Oil Content Extraction; Convolutional Neural Network; Model Integration; Comprehensive Image-Processing Techniques
Algorithm and Image Processing
GROUND OBJECT CLASSIFICATION ALGORITHM BASED ON ZHUHAI-1 SATELLITE
HYPERSPECTRAL IMAGE
Yun Zhou1, Jiangying Wang2 and Miaofen Huang3*
1 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China, Email: 1282854954@qq.com
2 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China, Email: 1361748619@qq.com
3 Guangdong Ocean University, No. 1, Haida Road, Mazhang District, Zhanjiang, 524088, China,* Corresponding author Email: hmf808@163.com
ABSTRACT
With the increasing resolution of satellite and aerial remote sensing images, much more useful spectral and spatial information can be obtained from hyperspectral images than before. This paper proposed a feature classification method based on machine learning and an integration model to classify ground objects in Zhuhai-1 OHS hyperspectral satellite images precisely and robustly. Firstly, principal component analysis was used to reduce the dimension of the data. Secondly, the classification model was established by the combination of undersampling, binary classification model and multi classification model. Then, random forest, AdaBoost and neural network classification algorithms were used to train the training data. Finally, based on the idea of voting method, the classification models were integrated into a new classification model, and comparisons were made between the classification of these single algorithms and the integration model. The principal component analysis on the hyperspectral images shows that the first two bands of these images almost contain 95% information. Therefore, the classification on the first two bands not only can maintain the accuracy, but also can reduce the amount of data for image classification processing and saving classification time effectively. Among the three single classification algorithms, the best one is the random forest classification model with an accuracy of 0.656 and a kappa value of 0.472. While the integration model can give a better classification result than the three-single classification algorithms, with an accuracy of 0.660 and the kappa value of 0.481, the integration model is in fact can improve the accuracy of the classification on Zhuhai-1 satellite hyperspectral images.
Keywords: Model Integration; Random Forests; Adaboost; Neural Networks
Algorithm and Image Processing
A NOVEL METHOD FOR MEASUREMENT OF
ORIENTATION OF ROCK JOINTS FROM POINT CLOUD BY FACET AMALGAMATION APPROACH
Anthony C.T. So, W.K. Leung and Jeffrey C.F. Wong
Geotechnical Engineering Office, Civil Engineering and Development Department, 101 Princess Margaret Road, Homantin, Kowloon, Hong Kong
Email: actso@cedd.gov.hk
ABSTRACT
The Geotechnical Engineering Office (GEO) of Civil Engineering and Development Department (CEDD) of the Government of the Hong Kong Special Administrative Region has been employing various remote sensing methods to enhance the engineering geological input in supporting geotechnical studies. For slope mapping, the use of conventional manual mapping method to measure joint orientations (i.e. dip angles and dip directions) has been found to be expensive, time-consuming and environmentally unfriendly. In addition, there may be constraints to access to remote sites and mapping of rock slope on site may pose safety hazards to the field personnel. To improve effectiveness and efficiency, many studies have been carried out to reduce the amount of field work by remote sensing techniques, and the focus was mainly on computer-aid generation of individual joint planes from point cloud captured either by laser scanning or photogrammetry. In view that such an approach often requires good understandings of the conditions of rock joints such as variations in orientations in different parts of a joint planes due to waviness and unevenness. To this end, delineation of joint plane requires setting of parameters by a trial and error approach which is not technically desirable. To overcome this problem, we have developed a new approach which is entirely different from the aforementioned computer-aid approach. Instead of attempting to form joint planes from point cloud, we first used point cloud to generate a 3-D triangular mesh to model the slope face. By measuring the dips and dip directions of all triangles (facets) of the mesh, a stereoplot of all facets was generated. Statistically, the majority of facets should be able to represent the overall orientations of all measurable discontinuities on the slope, and the facets due to waviness and unevenness in random directions should be in minority. To this end, by amalgamation of the orientations of facets, the major joint sets of the slope could be analysed. An assessment of different algorithms using K-d Tree, Fast Marching and this new approach have been performed. The results show that this novel method is more effective and efficient in identifying rock joints as compared to other algorithms with less requirements and more tolerance in parameter setting. This new approach has been proven to be an easy-to-use and user-friendly method which can greatly facilitate rock slope stability analysis.
Keywords: Rock Slopes, Joint Planes, Point Cloud, Facets, Dips, Dip Directions,
Algorithm and Image Processing
AN ANALYSIS ON THE OPTIMAL SEGMENTATION OF VHR SATELLITE RGB IMAGES
Jiyoon Moon1 and Kwangjae Lee2
1 Korea Aerospace Research Institute (KARI), 169-84, Gwahak-ro, Yuseong-gu, Daejeon 34133, S. Korea, Email: jymoon@kari.re.kr
2 Korea Aerospace Research Institute (KARI), 169-84, Gwahak-ro, Yuseong-gu, Daejeon 34133, S. Korea, Email: kjlee@kari.re.kr
ABSTRACT
With the development of satellites in the form of high-resolution and constellation satellites, it is necessary to mosaic and analyze dozens of images due to the narrower swath of high-resolution satellites compared to that of medium- or low-resolution satellites. As such, various studies using mosaic images are needed, and since these mosaic images have distorted spectral characteristics, there have been many restrictions on their use.
Therefore, in this study, an analysis on the segmentation using only RGB bands was conducted to utilize the images with distorted spectral characteristics for segmentation and even classification.
According to the segmentation results, the optimal segmentation scale was 75 for the mosaic images of Korean Peninsula, and the combination of RGB and Vegetation index seems the most appropriately segment the image especially for the boundary of bare ground and buildings. However, this study is optimized for mosaic images of Korean Peninsula, and it is necessary to compare and analyze the research results using other images and the results applied to other regions. Therefore, further research using those segmentation results should be made to validate and promote the use of mosaic images.
Keywords: KOMPSAT, Satellite Images, Segmentation, RGB
Algorithm and Image Processing
MODELLING SOIL ORGANIC CARBON STOCKS UNDER COMMERCIAL FORESTRY IN KWAZULU-NATAL
SOUTH AFRICA USING TOPO-CLIMATE VARIABLES
Omosalewa Odebiri1*, John Odindi1, Onisimo Mutanga1 and Kabir Peerbhay1
1 School of Agriculture, Earth and Environmental Science University of KwaZulu-Natal, Department of Geography, South Africa.
Email: Odindi@ukzn.ac.za
ABSTRACT
Commercial forests (CFs) are expanding globally, offering great potential for absorbing carbon and mitigating climate change. Compared to natural forests (NFs), CFs landscapes are largely ignored in soil organic carbon (SOC) mapping and climate change mitigation. Specifically, the relationship between the controlling factors that include topo-climate variables and the distribution of SOC is still poorly understood. Consequently, this study sought to map SOC stock variability within CFs using topo- climatic variables and geospatial strategies. Eighty one soil samples and 31 topo-climate predictors were simulated for SOC. A backward elimination method and the Maximum Entropy (Maxent) algorithm were used for optimum variable selection (11 variables) and regression, respectively.
Results showed good accuracies for both training (area under the curve = 0.906) and test (area under the curve = 0.885) datasets, and provide an effective framework for SOC modelling within CFs; valuable for climate change mitigation.
Keywords: Commercial forests; Soil Organic Carbon; Topo-climate; Maxent
Algorithm and Image Processing
ATMOSPHERIC CORRECTION FOR INLAND WATERS USING ARTIFICIAL NEURAL NETWORKS
Oanh Thi La1, Chao-Hung Lin1*, Ha Nguyen Thi Thu2, Manh Van Nguyen1,3 and Bo-Yi Lin1
1 Department of Geomatics, National Cheng Kung University, Tainan 70101, Taiwan, Emails: p66087063@mail.ncku.edu.tw, https://orcid.org/0000-0001- 8126-8794, p68087019@mail.ncku.edu.tw, bo-yi@gs.ncku.edu.tw
2 Faculty of Geology, VNU University of Science, Ha Noi 100000, Vietnam, Email: hantt_kdc@vnu.edu.vn
3 Institute of Geography, Vietnam Academy of Science and Technology, Hanoi 100000, Vietnam
ABSTRACT
The retrieval of water remote-sensing reflectance is an important and fundamental step for water quality monitoring using satellite remote sensing techniques. The atmospheric effects are generally significant and complex, which make atmospheric corrections (ACs) difficult to accurately derive the remote-sensing reflectance. The radiative transfer model is considered a promising method in atmospheric correction. Nevertheless, the approach requires calculating a set of parameters using complicated models and formulas, including aerosol model, atmospheric conditions, and sensor geometric information. This leads to time-consuming and sometimes produces negative water remote sensing reflectance. With a revisit cycle of 16-day, free available, and a high resolution of 30 meters, Landsat 8 OLI imagery is widely utilized for water quality monitoring in inland waters. However, the sensor receives surface reflectance and atmospheric effects, including scattering caused by Rayleigh and aerosol and absorption caused by gas and aerosol. This study proposed an atmospheric correction method based on artificial neural networks for inland waters to retrieve the water remote sensing reflectance (Rrs) using Landsat 8 OLI imagery. The input data required for the neural network model consists of a training dataset and a testing dataset. The number of data for training is 262580 samples, and 71 samples for testing. The training dataset includes eight TOA spectral reflectance (from band 1 to band 9 except the Panchromatic band), three geometric angles data, that is, sensor zenith angle (VZA), sun zenith angle (SZA), and relative azimuth angle (RAA), and aerosol data (AOT). The iCOR Rrses are used as the labels of the training dataset. The in-situ Rrs data which was measured in the field campaigns in Vietnam in different lakes and on different Landsat 8 acquisition days using spectra radiometer are separated into two groups: one for training the neural network model and the remaining data is for testing the model. The top of atmospheric (TOA) reflectance can be converted from digital number (DNs) using metadata file attached in Landsat 8 collection 1 Level 1 product, while the geometric
Algorithm and Image Processing
data (AOT) is attached in Landsat 8 level 2 product. The iCOR Rrs is the Rrs data retrieved by using image correction method (iCOR), which is assessed as one of the best radiative transfer models for atmospheric correction for inland water bodies. Our proposed model includes three 3-dimensional convolution layers to extract the TOA spectral feature, five fully connected layers to predict the Rrs, and one output layer. Besides, the model also contains one target layer where the iCOR Rrs and in-situ Rrs working as label data. The output Rrs is in five bands in the visible and near-infrared region. The Keras Tuner function was used for model tuning to obtain the optimal hyperparameters, including the number of hidden layers, the number of neurons for each layer, the dropout rate for each layer, the learning rate, and the best epoch. The proposed AC model was then retrained with the optimal hyperparameters. The retrieved Rrs results were then validated with in-situ measurements and compared with existing atmospheric correction methods (including dark object subtraction (DOS), quick atmospheric correction (QUAC), atmospheric correction for OLI lite (ACOLITE), fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH), Landsat 8 surface reflectance code (LaSRC), and Image correction for atmospheric effects (iCOR). The testing data were classified into four trophic classes (including oligotrophic, mesotrophic, eutrophic, and hypereutrophic). The testing results show that the retrieved Rrs values closely match with in-situ measurements in all five bands in the visible and near- infared region. The result also reveals that the proposed AC model can avoid producing negative remote-sensing reflectance. Comparing to the six AC methods, the proposed AC model shows the best performance in all trophic levels; while the iCOR processor is more appropriate for eutrophic water, the ACOLITE and LaSRC can be used, but they often failed in the NIR region. In all trophic levels, the DOS and QUAC methods seem unsuitable for atmospheric correction in inland lakes because they produce uncertainties by only removing the haze effect in the atmosphere. The proposed AC model was further tested on lake Laguna, Philippines, and lake Barra Bonita, Brazil, located in the tropical region, bringing reliable results. The proposed method can address the negative remote sensing reflectance issue existing in some traditional atmospheric correction methods. This method is efficient and easy to use when the network is well trained. Therefore, the proposed AC model has the potential for further remote sensing applications in water quality monitoring. However, the proposed model could not estimate Rrs reasonably in hypereutrophic water in algae bloom conditions. Thus, the proposed model is recommended for atmospheric correction in water bodies that have the Chl-a concentration in the range (1.6, 395) mg/m3 or Secchi disk depth (SD) in the range (0.2, 5.1) meter.
Keywords: Atmospheric correction, inland waters, artificial neural networks, Landsat 8 OLI imagery
Algorithm and Image Processing
SPECKLE FILTERING AND PHYSICAL SCATTERING DECOMPOSITION FOR ALOS-2 PALSAR-2
POLARIMETRIC MOSAIC
Ken Yoong LEE1*, Chen Guang HOU1, Soo Chin LIEW1 and Leong Keong KWOH1 Centre for Remote Imaging, Sensing and Processing National University of Singapore
18 Kent Ridge Road, Singapore 119227
Email: {crslky, crshc, scliew, crsklk}@nus.edu.sg
ABSTRACT
This paper presents the speckle filtering and physical scattering decomposition for the ALOS-2 PALSAR-2 fully polarimetric mosaic. The processing steps include 1) multi-looking, 2) speckle filtering, 3) geocoding, 4) mosaicking, and 5) model-based polarimetric decomposition. Another variant of iterative bilateral filter, called gravitational filter, was assessed for polarimetric speckle filtering. Benchmarking against the refined Lee filter and the boxcar filter, the experimental results on the ALOS-2 PALSAR-2 polarimetric data confirmed the effectiveness of the gravitational filter in speckle reduction and image feature retention. Prior to the mosaicking, the geodetic coordinates of selected control points, which were given based on the Geodetic Reference System 1980, were first converted into the Universal Transverse Mercator map coordinates. Subsequently, the ALOS-2 PALSAR- 2 specklefiltered data were geocoded by using second-order polynomial equation and least squares method. For the polarimetric scattering decomposition, an iterative multistage four-component decomposition was applied to the ALOS-2 PALSAR-2 polarimetric mosaic. From the decomposition result, a large amount of negative power pixels over vegetated areas were reduced. The total number of the remaining negative power pixels was only 0.006%.
Keywords: ALOS-2 PALSAR-2, Geocoding, Mosaicking, Speckle Filtering, Physical Scattering Decomposition
Algorithm and Image Processing
OBJECT-BASED CLASSIFICATION USING MASK R-CNN AND CNN FROM VERY HIGH-RESOLUTION IMAGERY
Batbold Badamdorj1* and Bolorchuluun Chogsom1
1 National University of Mongolia, Ikh Surguuliin Gudamj-1, Ulaanbaatar, Mongolia, Email: batbold9909@gmail.com
ABSTRACT
With the development of modern technology, the resolution of remote sensing imagery increasing rapidly. As the resolution of remote sensing imagery increases, it becomes more difficult to classify urban land useland cover types and recognize complex patterns in urban areas using traditional pixel-based methods. Therefore, we combined Object-based Image Analysis and Deep Learning, to accurately classify very high-resolution imagery of urban areas in this study. Object-based Image Analysis is used as the main classifier, while Deep Learning algorithms, Mask R-CNN and CNN, are used as feature extractors. In this study, very high-resolution images were used to classify three urban scenes in Ulaanbaatar, Mongolia. The first scene was used as the training site, whereas the second and the third scenes were used as test sites. As a result, the final classification with ten classes of selected scenes in Ulaanbaatar was created. The overall accuracy of classification result was above 90% in each scene, including 92.69% in the training site, 91.76% in the first test site, and 93.04% in the second test site. Our result shows the combination of Object-based Image Analysis and Deep Learning increases accuracy of classification from very high-resolution imagery of urban areas.
Keywords: Object-based Image Analysis, Deep learning, Mask R-CNN, CNN
Algorithm and Image Processing
IDENTIFICATION OF PERSISTENT INFRARED EMITTERS IN ASIA WITH VIIRS NIGHTFIRE DATA: 2012-2020
Christopher D. Elvidge1, Mikhail Zhizhin1-2, Feng Chi Hsu1, Tilottama Ghosh1 and Tamara Sparks1
1 Earth Observation Group, Payne Institute for Public Policy, Colorado School of Mines, Email: celvidge@mines.edu
2 Russian Space Research Institute, Moscow, Russia
ABSTRACT
The Visible Infrared Imaging Radiometer Suite (VIIRS) collects nightly global data in near infrared (NIR), shortwave infrared (SWIR) and midwave infrared (MWIR) spectral bands, providing a unique capability to observe and characterize infrared emitters at night. The VIIRS nightfire (VNF) algorithm identifies infrared (IR) emitters in multiple spectral bands and calculates temperature, source area and radiant heat via Planck curve fitting. VNF data are produced nightly and extend from 2012 to present. The most common infrared emitter across Asia is biomass burning. Industrial IR emitters are hidden amongst the vast numbers of biomass burning detections.
Here we present a survey of persistent IR emitters in Asia. Having a catalog of known IR emitter sites make it possible to monitor the sites for use in economic forecasting and greenhouse gas emission inventories.
Keywords: VIIRS, Nightfire, Infrared Emitters, Flares, Shortwave Infrared