Aaron E. Maxwell

Kernel-based extreme learning machine for remote-sensing image classification

This letter evaluates the effectiveness of a new kernel-based extreme learning machine (ELM) algorithm for a land cover classification using both multi- and hyperspectral remote-sensing data. The results are compared with the most widely used algorithms – support vector machines (SVMs). The results are compared in terms of the ease of use (in terms of the number of user-defined parameters), classification accuracy and computation cost. A radial basis kernel function was used with both the SVM and the kernel-based extreme-learning machine algorithms to ensure compatibility in the comparison of the two algorithms. The results suggest that the new algorithm is similar to, or more accurate than, SVM in terms of classification accuracy, has notable lower computational cost and does not require the implementation of a multiclass strategy.

Comparison of NAIP orthophotography and RapidEye satellite imagery for mapping of mining and mine reclamation

National Agriculture Imagery Program (NAIP) orthophotography is a potentially useful data source for land cover classification in the United States due to its nationwide and generally cloud-free coverage, low cost to the public, frequent update interval, and high spatial resolution. Nevertheless, there are challenges when working with NAIP imagery, especially regarding varying viewing geometry, radiometric normalization, and calibration. In this article, we compare NAIP orthophotography and RapidEye satellite imagery for high-resolution mapping of mining and mine reclamation within a mountaintop coal surface mine in the southern coalfields of West Virginia, USA. Two classification algorithms, support vector machines and random forests, were used to classify both data sets. Compared to the RapidEye classification, the NAIP classification resulted in lower overall accuracy and kappa and higher allocation disagreement and quantity disagreement. However, the accuracy of the NAIP classification was improved by reducing the number of classes mapped, using the near-infrared band, using textural measures and feature selection, and reducing the spatial resolution slightly by pixel aggregation or by applying a Gaussian low-pass filter. With such strategies, NAIP data can be a potential alternative to RapidEye satellite data for classification of surface mine land cover.

Scenario analysis predicts context-dependent stream response to landuse change in a heavily mined central Appalachian watershed

Abstract.  Scenario analysis has the potential to improve management of aquatic systems throughout the Mountaintop Removal–Valley Fill mining (MTR–VF) region of central Appalachia. However, the extent to which surface mining interacts with other landuse stressors (i.e., cumulative effects) is unclear, and this limits our ability to predict the effects of new mines on physical, chemical, and biological conditions downstream. We tested for additive and interactive effects of landuse change (surface mining, deep mining, and residential development) on water quality (specific conductance and Se), habitat quality, and benthic macroinvertebrates via a uniquely designed watershed-scale assessment of the Coal River, West Virginia (USA). We derived equations for predicting in-stream response to landscape changes and predicted the outcome of a realistic future scenario involving development of 15 permitted mines. Elevated Se concentrations were directly correlated with incremental increases in surface-mining extent. Surface mining, deep mining, and residential development had additive effects on elevated specific conductance and reduced biological condition. We found evidence of a positive interactive effect (stressor antagonism) of deep mining and residential development on biological condition, presumably caused by stream-flow augmentation from deep mines. Landscape context influenced predicted impacts from construction of 15 new mines because of additive and interactive effects of landuse change. New surface mines increased the number of receiving streams exceeding chemical and biological criteria, but a greater proportion of receiving streams exceeded chemical and biological criteria at equivalent levels of new mine development when pre-existing stressors were present. When surface mining was the only stressor, ≥30 or 40% increases in surface mining caused 100% of streams to exceed chemical or biological standards, respectively, whereas in streams stressed by deep mining and residential development, ≥10% additional surface mining caused 100% of streams to exceed chemical and biological standards. Continued progress in this area will require a better understanding of how landuse change affects aquatic systems in the rest of the MTR–VF mining region, where watershed-to-watershed variation in landuse patterns probably causes variability in ecological response.

Assessing machine-learning algorithms and image-and lidar-derived variables for GEOBIA classification of mining and mine reclamation

This study investigates machine-learning algorithms and measures derived from RapidEye satellite imagery and light detection and ranging (lidar) data for geographic object-based image analysis classification of mining and mine reclamation. Support vector machines, random forests, and boosted classification and regression trees classification algorithms were assessed and compared with the k-nearest neighbour (k-NN) classifier. For geographic object-based image analysis classification of mine landscapes, the use of disparate data (i.e. lidar data) improved overall accuracy, whereas the use of complex, object-oriented variables such as object geometry measures, first-order texture, and second-order texture from the grey-level co-occurrence matrix decreased or did not improve the classification accuracy. Support vector machines generally outperformed k-NN and the ensemble tree classifiers when only using the band means. With the incorporation of lidar-descriptive statistics, all four algorithms provided statistically comparable accuracies. K-NN suffered reduced classification accuracy with high-dimensional feature spaces, suggesting that a more complex machine-learning algorithm may be more appropriate when a large number of predictor variables are used.

Combining RapidEye Satellite Imagery and Lidar for Mapping of Mining and Mine Reclamation

The combination of RapidEye satellite imagery and light detection and ranging (lidar) derivatives was assessed for mapping land-cover within a mountaintop coal surface mine complex in the southern coalfields of West Virginia, USA. Support vector machines (SVM), random forests (RF), and boosted classification and regression trees (CART) algorithms were used. Incorporation of the lidar-derived data increased map accuracy in comparison to using only the five imagery bands, and SVM generally produced a more accurate classification than the ensemble tree algorithms based on overall map accuracy, Kappa statistics, allocation disagreement, quantity disagreement, and McNemar’s test of statistical significance. Based on measures of predictor variable importance within the ensemble tree classifiers, the normalized digital surface model (nDSM) was found to be more useful than first return intensity data for differentiating the classes. Introduction Commercial satellite imagery such as Ikonos, GeoEye, and RapidEye provide high spatial resolution but low spectral resolution compared to sensors such as Landsat Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), or Moderate Resolution Imaging Spectrometer (MODIS) (Warner et al., 2009). Although high spatial resolution can yield fine detail for land-cover and vegetative mapping, classification is complicated by the increased spatial resolution and decreased spectral resolution. Fine spatial resolution tends to generate high internal variability within land-cover classes, which can lead to decreases in classification accuracy (Townshend, 1981; Cushnie, 1987; Townshend, 1992; Baker et al., 2013). This research investigated a potential means to enhance classification accuracy by combining high-resolution commercial satellite imagery with light detection and ranging (lidar) data. The analysis focused on mapping land-cover classes in a mountaintop coal surface mine complex in the southern coalfields of the eastern United States. Because surface mine complexes experience rapid change due to human disturbance and reclamation, they are particularly good examples of disturbed landscapes. Although this research focuses on mapping land-cover within a mountaintop coal mine, the challenges in mapping mining landscapes are typical of other disturbances, such as timber harvesting, urban sprawl, etc. This work adds to prior remote sensing of surface mines research by investigating information gained by combining lidar and commercial satellite data for mapping land-cover (Cowen et al., 2000). This research had two components. First, we assessed lidar-derived inputs as predictor variables when combined with commercial satellite imagery to enhance land-cover mapping. Second, we compared three machine learning algorithms for the classification: support vector machines (SVM), random forests (RF), and boosted classification and regression trees (CART). The image data consisted of commercial RapidEye imagery. Lidar-derived predictor variables included the normalized digital surface model (nDSM) generated by subtracting ground return data from the first return data, first return intensity data, and the first return intensity range within raster grid cells.

Implementation of machine-learning classification in remote sensing: an applied review

ABSTRACT Machine learning offers the potential for effective and efficient classification of remotely sensed imagery. The strengths of machine learning include the capacity to handle data of high dimensionality and to map classes with very complex characteristics. Nevertheless, implementing a machine-learning classification is not straightforward, and the literature provides conflicting advice regarding many key issues. This article therefore provides an overview of machine learning from an applied perspective. We focus on the relatively mature methods of support vector machines, single decision trees (DTs), Random Forests, boosted DTs, artificial neural networks, and k-nearest neighbours (k-NN). Issues considered include the choice of algorithm, training data requirements, user-defined parameter selection and optimization, feature space impacts and reduction, and computational costs. We illustrate these issues through applying machine-learning classification to two publically available remotely sensed data sets.

Landscape-based cumulative effects models for predicting stream response to mountaintop mining in multistressor Appalachian watersheds

We conducted a survey of 170 streams distributed throughout the mountaintop-mining region of West Virginia (USA) and linked stream data to a temporally consistent and comprehensive land-cover data set. We then applied a generalized linear modeling framework and constructed cumulative effects models capable of predicting in-stream response to future surface-mine development within the context of other landuse activities. Predictive models provided precise estimates of specific conductance (model R2 ≤ 0.77 and cross-validated R2 ≤ 0.74), Se (0.74 and 0.70), and benthic macroinvertebrate community composition (0.72 and 0.65). Deletion tests supported the conclusion that stream degradation across the region is the result of complex, but predictable, additive and interactive effects of surface mining, underground mining, and residential development. Furthermore, we found that as stressors other than surface mining are factored out completely, the surface-mining level that results in exceedance of the 300 µS/cm conductivity benchmark increased from 4.4% in the presence of other stressors to 16.6% when only surface mining was present. Last, extrapolating model results to all unsampled stream segments in the region (n = 26,135), we predicted high levels of chemical (33%) and biological (67%) impairment to streams on the current landscape. Of this total impairment, however, <25% could be attributed to surface mining alone. These results underscore the importance of multistressor landuse models for reliable predictions of stream conditions, and the difficulty of interpreting correlations between surface mining and stream impairment without fully accounting for other landuse activities.

Differentiating mine-reclaimed grasslands from spectrally similar land cover using terrain variables and object-based machine learning classification

Incorporating ancillary, non-spectral data may improve the separability of land use/land cover classes. This study investigates the use of multi-temporal digital terrain data combined with aerial National Agriculture Imagery Program imagery for differentiating mine-reclaimed grasslands from non-mining grasslands across a broad region (6085 km2). The terrain data were derived from historical digital hypsography and a recent light detection and ranging data set. A geographic object-based image analysis (GEOBIA) approach, combined with two machine learning algorithms, Random Forests and Support Vector Machines, was used because these methods facilitate the use of ancillary data in classification. The results suggest that mine-reclaimed grasslands can be mapped accurately, with user’s and producer’s accuracies above 80%, due to a distinctive topographic signature in comparison with other spectrally similar grasslands within this landscape. The use of multi-temporal digital elevation model data and pre-mining terrain data only generally provided statistically significant increased classification accuracy in comparison with post-mining terrain data. Elevation change data were of value, and terrain shape variables generally improved the classification. GEOBIA and machine learning algorithms were useful in exploiting these non-spectral data, as data gridded at variable cell sizes can be summarized at the scale of image objects, allowing complex interactions between predictor variables to be characterized.

Predicting Palustrine Wetland Probability Using Random Forest Machine Learning and Digital Elevation Data-Derived Terrain Variables

Abstract The probability of palustrine wetland occurrence in the state of West Virginia, USA, was mapped based on topographic variables and using random forests ( rf ) machine learning. Models were developed for both selected ecological subregions and the entire state. The models were first trained using pixels randomly selected from the United States National Wetland Inventory ( nwi ) dataset and were tested using a separate random subset from the nwi and a database of wetlands not found in the nwi provided by the West Virginia Division of Natural Resources ( wvdnr ). The models produced area under the curve ( auc ) values in excess of 0.90, and as high as 0.998. Models developed in one ecological subregion of the state produced significantly different auc values when applied to other subregions, indicating that the topographical models should be extrapolated to new physiographic regions with caution. Several previously unexplored dem -derived terrain variables were found to be of value, including distance from water bodies, roughness, and dissection. Non- nwi wetlands were mapped with an auc value of 0.956, indicating that the probability maps may be useful for finding potential palustrine wetlands not found in the nwi .

Modeling Critical Forest Habitat in the Southern Coal Fields of West Virginia

Throughout the Central Appalachians of the United States resource extraction primarily from coal mining has contributed to the majority of the forest conversion to barren and reclaimed pasture and grass. The loss of forests in this ecoregion is significantly impacting biodiversity at a regional scale. Since not all forest stands provide equal levels of ecological functions, it is critical to identify and map existing forested resources by the benefits that accrue from their unique spatial patterns, watershed drainage, and landscape positions. We utilized spatial analysis and remote sensing techniques to define critical forest characteristics. The characteristics were defined by applying a forest fragmentation model utilizing morphological image analysis, defining headwater catchments at a 1 : 24,000 scale, and deriving ecological land units (ELUs) from elevation data. Once critical forest values were calculated, it was possible to identify clusters of critical stands using spatial statistics. This spatially explicit method for modeling forest habitat could be implemented as a tool for assessing the impact of resource extraction and aid in the conservation of critical forest habitat throughout a landscape.