Jagannath Aryal

Object-Based Classification of Ikonos Imagery for Mapping Large-Scale Vegetation Communities in Urban Areas

Effective assessment of biodiversity in cities requires detailed vegetation maps. To date, most remote sensing of urban vegetation has focused on thematically coarse land cover products. Detailed habitat maps are created by manual interpretation of aerial photographs, but this is time consuming and costly at large scale. To address this issue, we tested the effectiveness of object-based classifications that use automated image segmentation to extract meaningful ground features from imagery. We applied these techniques to very high resolution multispectral Ikonos images to produce vegetation community maps in Dunedin City, New Zealand. An Ikonos image was orthorectified and a multi-scale segmentation algorithm used to produce a hierarchical network of image objects. The upper level included four coarse strata: industrial/commercial (commercial buildings), residential (houses and backyard private gardens), vegetation (vegetation patches larger than 0.8/1ha), and water. We focused on the vegetation stratum that was segmented at more detailed level to extract and classify fifteen classes of vegetation communities. The first classification yielded a moderate overall classification accuracy (64%, κ = 0.52), which led us to consider a simplified classification with ten vegetation classes. The overall classification accuracy from the simplified classification was 77% with a κ value close to the excellent range (κ = 0.74). These results compared favourably with similar studies in other environments. We conclude that this approach does not provide maps as detailed as those produced by manually interpreting aerial photographs, but it can still extract ecologically significant classes. It is an efficient way to generate accurate and detailed maps in significantly shorter time. The final map accuracy could be improved by integrating segmentation, automated and manual classification in the mapping process, especially when considering important vegetation classes with limited spectral contrast.

Predictive habitat modelling to estimate petrel breeding colony sizes: Sooty shearwaters (Puffinus griseus) and mottled petrels (Pterodroma inexpectata) on Whenua Hou Island

Abstract Between 2001 and 2006, we systematically sampled the entire coast of Whenua Hou, a rugged offshore island in southern New Zealand, to estimate the population densities of sooty shearwaters (Puffinus griseus) and mottled petrels (Pterodroma inexpectata) by counting the entrances of breeding burrows. A two‐step regression modelling process using binomial errors was used to predict the presence of a colony, and a normal general linear model was used to predict the density of entrances within colonies. Aerial photography, GIS and a Digital Elevation Model were used to extract relevant habitat and location variables, and a combination of both regression models was used to predict the density of breeding burrows within each 5.32 m2 pixel on the island. This complex GIS and habitat prediction modelling approach gave population estimates very similar to a more traditional simple area extrapolation method and gave no improvement in precision. However, correction for the slope of the land increased our simple area estimates of population size by 11%. We estimate populations of sooty shearwater and mottled petrel breeding pairs at 173 000 (162 000–190 000) and 160 000 (123 000–197 000) respectively. Based on this number of breeding pairs, we calculate that Whenua Hou supports a total population of 868 000 (554 000–1 270 000) sooty shearwaters. Our estimate of the total mottled petrel population 202 000 pairs (162 000–242 000) is comparable with the only published estimate, but could be an underestimate because mottled petrels are sometimes found in large burrows. More research for robust estimation of population trends is needed to assess the conservation status of mottled petrels.

Big data integration shows Australian bush-fire frequency is increasing significantly.

Increasing Australian bush-fire frequencies over the last decade has indicated a major climatic change in coming future. Understanding such climatic change for Australian bush-fire is limited and there is an urgent need of scientific research, which is capable enough to contribute to Australian society. Frequency of bush-fire carries information on spatial, temporal and climatic aspects of bush-fire events and provides contextual information to model various climate data for accurately predicting future bush-fire hot spots. In this study, we develop an ensemble method based on a two-layered machine learning model to establish relationship between fire incidence and climatic data. In a 336 week data trial, we demonstrate that the model provides highly accurate bush-fire incidence hot-spot estimation (91% global accuracy) from the weekly climatic surfaces. Our analysis also indicates that Australian weekly bush-fire frequencies increased by 40% over the last 5 years, particularly during summer months, implicating a serious climatic shift.

Comparing GIS-based support vector machine kernel functions for landslide susceptibility mapping

This study compares the predictive performance of GIS-based landslide susceptibility mapping (LSM) using four different kernel functions in support vector machines (SVMs). Nine possible causal criteria were considered based on earlier similar studies for an area in the eastern part of the Khuzestan province of southern Iran. Different models and the resulting landslide susceptibility maps were created using information on known landslide events from a landslide inventory dataset. The models were trained using landslide inventory dataset. A two-step accuracy assessment was implemented to validate the results and to compare the capability of each function. The radial basis function was identified as the most efficient kernel function for LSM with the resulting landslide susceptibility map showing the highest predictive accuracy, followed by the polynomial kernel function. According to the obtained results, it concluded that using SVMs can generally be considered to be an effective method for LSM while it demands careful consideration of kernel function. The results of the present research will also assist other researchers to select the best SVM kernel function to use for LSM.

Fuzzy Shannon Entropy: A Hybrid GIS-Based Landslide Susceptibility Mapping Method

Assessing Landslide Susceptibility Mapping (LSM) contributes to reducing the risk of living with landslides. Handling the vagueness associated with LSM is a challenging task. Here we show the application of hybrid GIS-based LSM. The hybrid approach embraces fuzzy membership functions (FMFs) in combination with Shannon entropy, a well-known information theory-based method. Nine landslide-related criteria, along with an inventory of landslides containing 108 recent and historic landslide points, are used to prepare a susceptibility map. A random split into training (≈70%) and testing (≈30%) samples are used for training and validation of the LSM model. The study area—Izeh—is located in the Khuzestan province of Iran, a highly susceptible landslide zone. The performance of the hybrid method is evaluated using receiver operating characteristics (ROC) curves in combination with area under the curve (AUC). The performance of the proposed hybrid method with AUC of 0.934 is superior to multi-criteria evaluation approaches using a subjective scheme in this research in comparison with a previous study using the same dataset through extended fuzzy multi-criteria evaluation with AUC value of 0.894, and was built on the basis of decision makers’ evaluation in the same study area.

Recommending environmental knowledge as linked open data cloud using semantic machine learning

Large scale environmental knowledge integration and development of a knowledge recommendation system for the Linked Open Data Cloud using semantic machine learning approach was the main mission of this research. This study considered five different environmental big data sources including SILO, AWAP, ASRIS, MODIS and CosmOz complementary for knowledge integration. Unsupervised clustering techniques based on principal component analysis (PCA) and Fuzzy-C-Means (FCM) and Self-organizing map (SOM) clustering was used to learn the extracted features and to create a 2D map based dynamic knowledge recommendation system. Knowledge was stored in a triplestore using triples format (subject, predicate, and object) along with the complete meta-data provenance information. The Resource Description Framework (RDF) representation made i-EKbase very flexible to integrate with the Linked Open Data (LOD) cloud. The developed Intelligent Environmental Knowledgebase (i-EKbase) could be used for any environmental decision support application.

Use of the Bradley-Terry model to quantify association in remotely sensed images

Thematic maps prepared from remotely sensed images require a statistical accuracy assessment. For this purpose, the /spl kappa/-statistic is often used. This statistic does not distinguish between whether one unit is classified as another, or vice versa. In this paper, the Bradley-Terry (BT) model is applied for accuracy assessment. This model compares categories pairwise. The probability of one class over another class is estimated as well as the expected values of class pixels. The study is illustrated with an Advanced Spaceborne Thermal Emission and Reflection Radiometer image from the Netherlands, to which a maximum-likelihood classification with the Euclidean distance is applied. An error matrix is generated using an IKONOS image from the same area as ground truth. It is shown to which degree the BT model extends the /spl kappa/-statistic. A comparison with the Mahalanobis distance is made. Standardization is carried out to overcome problems emerging from the fact that a common BT model does not include the number of correctly classified pixels. The study shows how the BT model serves as an alternative to the usual /spl kappa/-statistic.

Near-Real Time Detection of Beetle Infestation in Pine Forests Using MODIS Data

This paper considers near-real time detection of beetle infestation in North American pine forests using MODIS 8-days 500 m data. Two methods are considered, both using a single time series for detection of beetle infestation by analyzing the statistics of the trend component of the signal. The first method estimates the trend component of the vegetation index time series by fitting an underlying triply modulated cosine model over a sliding window, using nonlinear least squares (NLS), and the second method uses a T-point moving average finite impulse response (FIR) filter. Both the methods perform well and show similar performance on simulated datasets. The methods are also tested on many difference and ratio-indices of a real-world dataset with change and no-change examples taken from the Rocky Mountain region of the United States and of British Columbia in Canada. The results suggest that both the methods detect beetle infestation reliably in almost all the vegetation index datasets. However, the model-based method (NLS-based) performs better in terms of the detection delay. Red Green Index (RGI), when used with the model-based method, provides the best tradeoff between the detection delay and accuracy. Furthermore, 90%, 50%, and 25% cross-validations are also performed for the threshold selection on RGI dataset, and it is shown that the selected threshold works well on the test data. In the end, it is also shown that the model-based method outperforms a recently published method for near-real time disturbance detection in MODIS data, in both accuracy and detection delay.

Deep cognitive imaging systems enable estimation of continental-scale fire incidence from climate data

Unplanned fire is a major control on the nature of terrestrial ecosystems and causes substantial losses of life and property. Given the substantial influence of climatic conditions on fire incidence, climate change is expected to substantially change fire regimes in many parts of the world. We wished to determine whether it was possible to develop a deep neural network process for accurately estimating continental fire incidence from publicly available climate data. We show that deep recurrent Elman neural network was the best performed out of ten artificial neural networks (ANN) based cognitive imaging systems for determining the relationship between fire incidence and climate. In a decennium data experiment using this ANN we show that it is possible to develop highly accurate estimations of fire incidence from monthly climatic data surfaces. Our estimations for the continent of Australia had over 90% global accuracy and a very low level of false negatives. The technique is thus appropriate for use in estimating the spatial consequences of climate scenarios on the monthly incidence of wildfire at the landscape scale.

A Statistical Framework for Near-Real Time Detection of Beetle Infestation in Pine Forests Using MODIS Data

Beetle infestations have caused significant damage to the pine forest in North America. Early detection of beetle infestation in near real time is crucial, in order to take appropriate steps to control the damage. In this letter, we consider near-real-time detection of beetle infestation in North American pine forests using high temporal resolution and coarse spatial resolution MODIS (eight-day 500-m) satellite data. We show that the parameter sequence of a stationary vegetation index time series, which is derived by fitting an underlying triply modulated cosine model over a sliding window using nonlinear least squares, resembles a martingale sequence. The advantage of such properties of the parameter sequence is that standard martingale central limit theorem and well-known Gaussian distribution statistics can be effectively used to detect any nonstationarity in the vegetation index time series with high accuracy. The proposed method exploits these properties of the parameter time series and, hence, does not require threshold tuning. The threshold is selected based on a well-documented procedure of z-value selection from the table of Gaussian distribution, depending upon the percentage of the distribution considered as outlier. The proposed framework is tested on different vegetation index data sets derived from MODIS eight-day 500-m image time series of beetle infestations in North America. The results show that the proposed framework can detect nonstationarities in the vegetation index time series accurately and performs the best on red-green index.