Multispectral and Texture Feature Application in Image-Object Analysis of Summer Vegetation in Eastern Tajikistan Pamirs

Combining Object-Based Machine Learning with Long-Term Time-Series Analysis for Informal Settlement Identification

Informal settlement mapping is essential for planning, as well as resource and utility management. Developing efficient ways of determining the properties of informal settlements (when, where, and who) is critical for upgrading services and planning. Remote sensing data are increasingly used to understand built environments. In this study, we combine two sources of data, very-high-resolution imagery and time-series Landsat data, to identify and describe informal settlements. The indicators characterising informal settlements were grouped into four different spatial and temporal levels: environment, settlement, object and time. These indicators were then used in an object-based machine learning (ML) workflow to identify informal settlements. The proposed method had a 95% overall accuracy at mapping informal settlements. Among the spatial and temporal levels examined, the contribution of the settlement level indicators was most significant in the ML model, followed by the object-level indicators. Whilst the temporal level did not contribute greatly to the classification of informal settlements, it provided a way of understanding when the settlements were formed. The adaptation of this method would allow the combination of a wide-ranging and diverse group of indicators in a comprehensive ML framework.

Decision-level fusion of Sentinel-1 SAR and Landsat 8 OLI texture features for crop discrimination and classification: case of Masvingo, Zimbabwe

Radar imagery have few polarization bands which can limit the ability to do traditional digital classification. Harmonization of Sentinel-1 and Landsat 8 data despite having complementary texture information can be a challenge. The objectives of this paper are to explore texture features derived from Landsat 8 OLI and dual-polarized Sentinel-1 SAR speckle filtered and unfiltered backscatter, to aggregate classification results using Decision-Level Fusion (DLF), and to evaluate the performance of decision-level fused maps. Gray Level Co-occurrence Matrix (GLCM) is employed to derive sets of seven texture features for Landsat 8 bands and VV + VH backscatter using 5 × 5, 7 × 7, 9 × 9, and 11 × 11 window sizes. Each texture feature is stacked with a respective source image and classified using Support Vector Machine (SVM). Classified maps from the best three performers from both speckle filtered and unfiltered are aggregated with classified maps from Landsat 8 using plurality voting algorithm and compared using Z-test. Results indicate an overall classification accuracy of 96.02% from DLF images of Landsat and non-speckle filtered maps, whereas Landsat and speckle filtered achieved 94.69%. The best texture information are derived from the blue band followed by the red band, whereas speckle unfiltered textures performed better than speckle filtered textures. We conclude that integration of Landsat 8 and Sentinel-1, either speckle filtered or unfiltered, improves crop classification and speckles do not have statistically significant effects (p = 0.1208).

A Novel Feature Extension Method for the Forest Disaster Monitoring Using Multispectral Data

Remote sensing images classification is the key technology for monitoring forest changes. Texture features have been demonstrated to have better effectiveness than spectral features in the improvement of the classification accuracy. The accuracy of extracting texture information by window-based method depends on the choice of the window size. Moreover, the size should ideally match the spatial scale of the object or class under consideration. However, most of the existing texture feature extraction methods are all based on a single window and do not adequately consider the scale of different objects. Our first proposition is to use a composite window for extracting texture features, which is a small window surrounded by a larger window. Our second proposition is to reinforce the performance of the trained ensemble classifier by training it using only the most important features. Considering the advantages of random forest classifier, such as fast training speed and few parameters, these features feed this classifier. Measures of feature importance are estimated along with the growth of the base classifiers, here decision trees. We aim to classify each pixel of the forest images disturbed by hurricanes and fires in three classes, damaged, not damaged, or unknown, as this could be used to compute time-dependent aggregates. In this study, two research areas—Nezer Forest in France and Blue Mountain Forest in Australia—are utilized to validating the effectiveness of the proposed method. Numerical simulations show increased performance and improved monitoring ability of forest disturbance when using these two propositions. When compared with the reference methods, the best increase of the overall accuracy obtained by the proposed algorithm is 4.77% and 2.96% on the Nezer forest data and Blue Mountain forest data, respectively.

Integration of Sentinel-1 and Sentinel-2 Data for Land Cover Mapping Using W-Net

In this paper, we present a new approach to the fusion of Sentinel 1 (S1) and Sentinel 2 (S2) data for land cover mapping. The proposed solution aims at improving methods based on Sentinel 2 data, that are unusable in case of cloud cover. This goal is achieved by using S1 data to generate S2-like segmentation maps to be used to integrate S2 acquisitions forbidden by cloud cover. In particular, we propose for the first time in remote sensing a multi-temporal W-Net approach for the segmentation of Interferometric Wide swath mode (IW) Sentinel-1 data collected along ascending/descending orbit to discriminate rice, water, and bare soil. The quantitative assessment of segmentation accuracy shows an improvement of 0.18 and 0.25 in terms of accuracy and F1-score by applying the proposed multi-temporal procedure with respect to the previous single-date approach. Advantages and disadvantages of the proposed W-Net based solution have been tested in the National Park of Albufera, Valencia, and we show a performance gain in terms of the classical metrics used in segmentation tasks and the computational time.

Modified shape index for object-based random forest image classification of agricultural systems using airborne hyperspectral datasets

This paper highlights the importance of optimized shape index for agricultural management system analysis that utilizes the contiguous bands of hyperspectral data to define the gradient of the spectral curve and improve image classification accuracy. Currently, a number of machine learning methods would resort to using averaged spectral information over wide bandwidths resulting in loss of crucial information available in those contiguous bands. The loss of information could mean a drop in the discriminative power when it comes to land cover classes with comparable spectral responses, as in the case of cultivated fields versus fallow lands. In this study, we proposed and tested three new optimized novel algorithms based on Moment Distance Index (MDI) that characterizes the whole shape of the spectral curve. The image classification tests conducted on two publicly available hyperspectral data sets (AVIRIS 1992 Indian Pine and HYDICE Washington DC Mall images) showed the robustness of the optimized algorithms in terms of classification accuracy. We achieved an overall accuracy of 98% and 99% for AVIRIS and HYDICE, respectively. The optimized indices were also time efficient as it avoided the process of band dimension reduction, such as those implemented by several well-known classifiers. Our results showed the potential of optimized shape indices, specifically the Moment Distance Ratio Right/Left (MDRRL), to discriminate between types of tillage (corn-min and corn-notill) and between grass/pasture and grass/trees, tree and grass under object-based random forest approach.

Object-Based Plastic-Mulched Landcover Extraction Using Integrated Sentinel-1 and Sentinel-2 Data

Plastic mulching on farmland has been increasing worldwide for decades due to its superior advantages for improving crop yields. Monitoring Plastic-Mulched Land-cover (PML) can provide essential information for making agricultural management decisions and reducing PML’s eco-environmental impacts. However, mapping PML with remote sensing data is still challenging and problematic due to its complicated and mixed characteristics. In this study, a new Object-Based Image Analysis (OBIA) approach has been proposed to investigate the potential for combined use of Sentinel-1 (S1) SAR and Sentinel-2 (S2) Multi-spectral data to extract PML. Based on the ESP2 tool (Estimation of Scale Parameter 2) and ED2 index (Euclidean Distance 2), the optimal Multi-Resolution Segmentation (MRS) result is chosen as the basis of following object-based classification. Spectral and backscattering features, index features and texture features from S1 and S2 are adopted in classifying PML and other land-cover types. Three machine-learning classifiers known as the—Classification and Regression Tree (CART), the Random Forest (RF) and the Support Vector Machine (SVM) are carried out and compared in this study. The best classification result with an overall accuracy of 94.34% is achieved by using spectral, backscattering, index and textural information from integrated S1 and S2 data with the SVM classifier. Texture information is demonstrated to contribute positively to PML classifications with SVM and RF classifiers. PML mapping using SAR information alone has been greatly improved by the object-based approach to an overall accuracy of 87.72%. By adding SAR data into optical data, the accuracy of object-based PML classifications has also been improved by 1–3%.

Mapping Population Distribution from High Resolution Remotely Sensed Imagery in a Data Poor Setting

Accurate mapping of population distribution is essential for policy-making, urban planning, administration, and risk management in hazardous areas. In some countries, however, population data is not collected on a regular basis and is rarely available at a high spatial resolution. In this study, we proposed an approach to estimate the absolute number of inhabitants at the neighborhood level, combining data obtained through field work with high resolution remote sensing. The approach was tested on Ngazidja Island (Union of the Comoros). A detailed survey of neighborhoods at the level of individual dwellings, showed that the average number of inhabitants per dwelling was significantly different between buildings characterized by a different roof type. Firstly, high spatial resolution remotely sensed imagery was used to define the location of individual buildings, and second to determine the roof type for each building, using an object-based classification approach. Knowing the location of individual houses and their roof type, the number of inhabitants was estimated at the neighborhood level using the data on house occupancy of the field survey. To correct for misclassification bias in roof type discrimination, an inverse calibration approach was applied. To assess the impact of variations in average dwelling occupancy between neighborhoods on model outcome, a measure of the degree of confidence of population estimates was calculated. Validation using the leave-one-out approach showed low model bias, and a relative error at the neighborhood level of 17%. With the increasing availability of high resolution remotely sensed data, population estimation methods combining data from field surveys with remote sensing, as proposed in this study, hold great promise for systematic mapping of population distribution in areas where reliable census data are not available on a regular basis.

Habitat assessment of Marco Polo sheep (Ovis ammon polii) in Eastern Tajikistan: Modeling the effects of climate change

Identifying the factors predicting the high-elevation suitable habitats of Central Asian argali wild sheep and how these suitable habitats are affected by the changing climate regimes could help address conservation and management efforts and identify future critical habitat for the species in eastern Tajikistan. This study used environmental niche models (ENMs) to map and compare potential present and future distributions of suitable environmental conditions for Marco Polo argali. Argali occurrence points were collected during field surveys conducted from 2009 to 2016. Our models showed that terrain ruggedness and annual mean temperature had strong correlations on argali distribution. We then used two greenhouse gas concentration trajectories (RCP 4.5 and RCP 8.5) for two future time periods (2050 and 2070) to model the impacts of climate change on Marco Polo argali habitat. Results indicated a decline of suitable habitat with majority of losses observed at lower elevations (3,300-4,300 m). Models that considered all variables (climatic and nonclimatic) predicted losses of present suitable areas of 60.6% (6,928 km2) and 63.2% (7,219 km2) by 2050 and 2070, respectively. Results also showed averaged habitat gains of 46.2% (6,106 km2) at much higher elevations (4,500-6,900 m) and that elevational shifts of habitat use could occur in the future. Our results could provide information for conservation planning for this near threatened species in the region.

Comparing Pixel- and Object-Based Approaches in Effectively Classifying Wetland-Dominated Landscapes

Wetland ecosystems straddle both terrestrial and aquatic habitats, performing many ecological functions directly and indirectly benefitting humans. However, global wetland losses are substantial. Satellite remote sensing and classification informs wise wetland management and monitoring. Both pixel- and object-based classification approaches using parametric and non-parametric algorithms may be effectively used in describing wetland structure and habitat, but which approach should one select? We conducted both pixel- and object-based image analyses (OBIA) using parametric (Iterative Self-Organizing Data Analysis Technique, ISODATA, and maximum likelihood, ML) and non-parametric (random forest, RF) approaches in the Barguzin Valley, a large wetland (~500 km²) in the Lake Baikal, Russia, drainage basin. Four Quickbird multispectral bands plus various spatial and spectral metrics (e.g., texture, Non-Differentiated Vegetation Index, slope, aspect, etc.) were analyzed using field-based regions of interest sampled to characterize an initial 18 ISODATA-based classes. Parsimoniously using a three-layer stack (Quickbird band 3, water ratio index (WRI), and mean texture) in the analyses resulted in the highest accuracy, 87.9% with pixel-based RF, followed by OBIA RF (segmentation scale 5, 84.6% overall accuracy), followed by pixel-based ML (83.9% overall accuracy). Increasing the predictors from three to five by adding Quickbird bands 2 and 4 decreased the pixel-based overall accuracy while increasing the OBIA RF accuracy to 90.4%. However, McNemar’s chi-square test confirmed no statistically significant difference in overall accuracy among the classifiers (pixel-based ML, RF, or object-based RF) for either the three- or five-layer analyses. Although potentially useful in some circumstances, the OBIA approach requires substantial resources and user input (such as segmentation scale selection—which was found to substantially affect overall accuracy). Hence, we conclude that pixel-based RF approaches are likely satisfactory for classifying wetland-dominated landscapes.

Summer and winter habitat suitability of Marco Polo argali in southeastern Tajikistan: A modeling approach

We modeled summer and winter habitat suitability of Marco Polo argali in the Pamir Mountains in southeastern Tajikistan using these statistical algorithms: Generalized Linear Model, Random Forest, Boosted Regression Tree, Maxent, and Multivariate Adaptive Regression Splines. Using sheep occurrence data collected from 2009 to 2015 and a set of selected habitat predictors, we produced summer and winter habitat suitability maps and determined the important habitat suitability predictors for both seasons. Our results demonstrated that argali selected proximity to riparian areas and greenness as the two most relevant variables for summer, and the degree of slope (gentler slopes between 0° to 20°) and Landsat temperature band for winter. The terrain roughness was also among the most important variables in summer and winter models. Aspect was only significant for winter habitat, with argali preferring south-facing mountain slopes. We evaluated various measures of model performance such as the Area Under the Curve (AUC) and the True Skill Statistic (TSS). Comparing the five algorithms, the AUC scored highest for Boosted Regression Tree in summer (AUC = 0.94) and winter model runs (AUC = 0.94). In contrast, Random Forest underperformed in both model runs.