Training Pixels Research Articles

Watershed-Based Attribute Profiles With Semantic Prior Knowledge for Remote Sensing Image Analysis

In this article, we develop a novel feature extraction method that combines two well-established mathematical morphology concepts: watersheds and morphological attribute profiles (APs). In order to extract spatial-spectral features from remote sensing data, APs were originally defined as sequences of filtering operators on inclusion trees, i.e., the max- and min-trees, computed from the input image. In this study, we extend the AP paradigm to the more general framework of hierarchical watersheds. Moreover, we explore the semantic knowledge provided by labeled training pixels during different phases of the watershed-AP construction, namely within the construction of hierarchical watersheds from the raw image and later within the filtering of the resulting hierarchy. We illustrate the relevance of the proposed method with two applications including land cover classification and building extraction using optical remote sensing images. Experimental results show that the new profiles outperform various existing features using two public datasets (Zurich and Vaihingen), thus providing another high potential feature extraction method within the AP family.

S3ANet: Spectral-spatial-scale attention network for end-to-end precise crop classification based on UAV-borne H2 imagery

High spatial and spectral resolution (H2) imagery collected by unmanned aerial vehicle (UAV) systems is an important data source for precise crop classification. Although this data source can provide us with abundant information about the crops of interest, it also introduces new challenges for the image processing. Specifically, the spectral similarities of green crops lead to small inter-class distances, and the severe intra-class spectral variability and high spatial heterogeneity in H2 imagery increases the difficulty of precise classification. In addition, the scales of the different crop plots can show great differences, which makes it difficult to determine the optimal patch size for deep learning based classification models. In this paper, a spectral-spatial-scale attention network (S3ANet) is proposed for H2 imagery based precise crop classification. In the proposed method, each channel, each pixel, and each scale perception of the feature map is adaptively weighted to relieve the intra-class spectral variability, the spatial heterogeneity, and the scale difference of the crop plots, respectively. Furthermore, the proposed S3ANet method introduces the additive angular margin loss function to further increase the inter-class distances between the different crops, and reduce the misclassification effect. S3ANet was verified using the public WHU-Hi UAV-borne hyperspectral dataset and the new WHU-Hi-JiaYu dataset, which is a dataset for precise rice classification that was built by the authors. In these experiments, the overall accuracy of the proposed S3ANet method all exceeds 96% under 50 training pixels per class, and it achieved significant improvement compared with some state-of-the-art hyperspectral images classifiers (such as SSRN, CNNCRF and FPGA, etc.). The code of S3ANet is available at http://rsidea.whu.edu.cn/resource_sharing.htm.

Qualitative Histopathological Classification of Primary Bone Tumors Using Deep Learning: A Pilot Study

BackgroundHistopathological diagnosis of bone tumors is challenging for pathologists. We aim to classify bone tumors histopathologically in terms of aggressiveness using deep learning (DL) and compare performance with pathologists.MethodsA total of 427 pathological slides of bone tumors were produced and scanned as whole slide imaging (WSI). Tumor area of WSI was annotated by pathologists and cropped into 716,838 image patches of 256 × 256 pixels for training. After six DL models were trained and validated in patch level, performance was evaluated on testing dataset for binary classification (benign vs. non-benign) and ternary classification (benign vs. intermediate vs. malignant) in patch-level and slide-level prediction. The performance of four pathologists with different experiences was compared to the best-performing models. The gradient-weighted class activation mapping was used to visualize patch’s important area.ResultsVGG-16 and Inception V3 performed better than other models in patch-level binary and ternary classification. For slide-level prediction, VGG-16 and Inception V3 had area under curve of 0.962 and 0.971 for binary classification and Cohen’s kappa score (CKS) of 0.731 and 0.802 for ternary classification. The senior pathologist had CKS of 0.685 comparable to both models (p = 0.688 and p = 0.287) while attending and junior pathologists showed lower CKS than the best model (each p < 0.05). Visualization showed that the DL model depended on pathological features to make predictions.ConclusionDL can effectively classify bone tumors histopathologically in terms of aggressiveness with performance similar to senior pathologists. Our results are promising and would help expedite the future application of DL-assisted histopathological diagnosis for bone tumors.

Stacked vector multi-source lithologic classification utilizing Machine Learning Algorithms: Data potentiality and dimensionality monitoring

Machine Learning Algorithms (MLAs) have recently introduced considerable lithologic mapping. Thus, this study scrutinizes the efficacy of Artificial Neural Network (ANN), Maximum Likelihood Classifier (MLC) and Support Vector Machine (SVM) over hybrid datasets including optical (Sentinel 2, ASTER, Landsat OLI and Earth-observing 1 Advanced Land Imager (ALI)), radar (Sentinel 1 and ALOS PALSAR), DEMs and their derivatives (Slope, and Aspect). The study aims to (1) monitor the effect of data dimensionality in enhancing categorization accuracy. (2) disclose the most efficient MLA and most powerful dataset in labeling rock units accurately. (3) highlight the impact of embedding topographical and radar data in lithologic classification. (4) outline the best relation between the number of training pixels and number of utilized bands, in delivering reliable allocation. To achieve these aims, we selected training and testing pixels meticulously, in concordance with a recently published geological map of the study area. We adopted a stacked vector approach for handling the implemented multi-sensor data. Results show that diversifying information sources raised the classification accuracy by approximately 10% for each classifier. SVM and MLC are much better than ANN. Slope is better than aspect and both are less qualified when compared to DEM. Sentinel 1 (C-band) and ALOS PALSAR (L-band) effects are not so different whatever the implemented polarization. Landsat OLI is less qualified in lithologic classification when compared to Sentinel 2, ASTER and ALI. The utilized training pixels should be at least 30N for (N) channels submitted to the classifiers.

Cloud Detection Using an Ensemble of Pixel-Based Machine Learning Models Incorporating Unsupervised Classification

Remote sensing imagery, such as that provided by the United States Geological Survey (USGS) Landsat satellites, has been widely used to study environmental protection, hazard analysis, and urban planning for decades. Clouds are a constant challenge for such imagery and, if not handled correctly, can cause a variety of issues for a wide range of remote sensing analyses. Typically, cloud mask algorithms use the entire image; in this study we present an ensemble of different pixel-based approaches to cloud pixel modeling. Based on four training subsets with a selection of different input features, 12 machine learning models were created. We evaluated these models using the cropped LC8-Biome cloud validation dataset. As a comparison, Fmask was also applied to the cropped scene Biome dataset. One goal of this research is to explore a machine learning modeling approach that uses as small a training data sample as possible but still provides an accurate model. Overall, the model trained on the sample subset (1.3% of the total training samples) that includes unsupervised Self-Organizing Map classification results as an input feature has the best performance. The approach achieves 98.57% overall accuracy, 1.18% cloud omission error, and 0.93% cloud commission error on the 88 cropped test images. By comparison to Fmask 4.0, this model improves the accuracy by 10.12% and reduces the cloud omission error by 6.39%. Furthermore, using an additional eight independent validation images that were not sampled in model training, the model trained on the second largest subset with an additional five features has the highest overall accuracy at 86.35%, with 12.48% cloud omission error and 7.96% cloud commission error. This model’s overall correctness increased by 3.26%, and the cloud omission error decreased by 1.28% compared to Fmask 4.0. The machine learning cloud classification models discussed in this paper could achieve very good performance utilizing only a small portion of the total training pixels available. We showed that a pixel-based cloud classification model, and that as each scene obviously has unique spectral characteristics, and having a small portion of example pixels from each of the sub-regions in a scene can improve the model accuracy significantly.

PERFORMANCE ASSESSMENT OF CLASSIFICATION ALGORITHMS FOR LANDUSE / LANDCOVER CHANGE USING SENTINEL 2 DATA – A CASE STUDY OF TIRUPPUR

Abstract. In the history of mankind, one of the vibrant geographical phenomena is urbanization. The urbanization process is characterized by the expansion of the city from the core to peripheral areas which includes economic development, social, political forces and population density. Very rapid urbanization in the highly populated country like India, which changes natural land cover into urban land use, which is unavoidable. However, the study region Tiruppur is known as the knitwear capital of India that induces urban development in the region which results in the modification of the natural land cover. For understating the interaction between the natural landscape and human activities, land use and land cover (LULC) is considered as the important indicator. Research on land-use and land cover changes using remote sensing technology has a long history to evident. The advancement in the Remote Sensing and GIS techniques provide the fine resolution of data sets to proceed. Sentinel-2B imagery was chosen for this study for two main reasons one is that compare to Landsat imagery it has a high spatial resolution of 10 m and its radiometry includes three vegetation red edge bands. These two characteristics make the Sentinel-2B data appealing for LULC mapping. Different types of classification algorithms have been used to perform land use and land cover mapping. The study aims to create land use and land cover classification by making a comparison between different algorithms in Tiruppur by using Sentinel-2B satellite imagery. The commonly known classification algorithms, K-means, IsoData, support vector machines (SVMs), and maximum likelihood (ML) classification are adopted for investigation. This is followed by the selection of training pixels from the remaining classes to perform and compare different supervised learning algorithms for the first- and second-level classification in terms of accuracy rates. Accuracy was assessed through metrics derived from an error matrix, but primarily overall accuracy and kappa coefficient was used in allocating algorithm hierarchy. Finally, after the comparison, the highly accurate algorithm was suggested for the mapping of urban areas. The highest overall accuracy and kappa coefficient was produced by support vector machine (SVM) is due to the algorithm’s relatively small number of complex decision boundaries. The results are helpful to understand the performance of the classification algorithm for the future studies.

A two-step mapping of irrigated corn with multi-temporal MODIS and Landsat analysis ready data

Timely and reliable information about irrigated croplands is important for crop water stress analysis and studies of water, energy, and food security. This study mapped irrigated and non-irrigated corn at 30 m resolution for the state of Nebraska using a two-step multi-temporal image classification of MODIS and Landsat Analysis Ready Data (ARD). Starting from the drought year of 2012, when there was a high contrast between irrigated and non-irrigated fields, we first conducted image classification using the 250 m MODIS multi-temporal NDVI data. Training pixels were automatically derived, based on counties with predominant irrigated and non-irrigated cornfields. The MODIS-derived irrigated vs. non-irrigated map was further spatially filtered to generate training data covering the entire Nebraska to support automated Landsat ARD classification, footprint-by-footprint. Three classification algorithms of multi-layer perceptron (MLP) neural network, Random Forest (RF), and Support Vector Machine (SVM) were implemented to classify all available Landsat ARD images within the growing season (i.e. May to November). Given the issues of scanline corrector (SLC) error and cloud contamination, the provisional Landsat-based classifications were finally gap-filled to generate a seamless state-wise irrigation map guided by decreasing cross-validation accuracy. Pixel-wise accuracy assessments showed similar overall accuracies of 89.6%, 89.3%, and 90.0% for MLP, RF, and SVM, respectively. They are 3-6% higher than a commonly used gap-filing procedure based on valid (cloud free) pixel count for growing season images. The estimated areas of irrigated corn from Landsat-based mapping were consistent with the 2012 USDA county level census data (R2 = 0.97 and RMSE = 37.70 km2). Using the 2012 Landsat-derived irrigation map and the USDA’s annual Cropland Data Layer as inputs, we further developed training data for annual irrigation mapping between 2013 and 2018. Pixel-wise assessment of the 2016 map showed reasonable overall accuracies of 78.4–79.6% for three classification algorithms. The annual maps yielded R2 of 0.94–0.98 and RMSE values of 37.70–57.62 km2 for various mapping years compared with USDA county statistics. These results suggest that our proposed two-step analytical method has a high potential for automated annual irrigation mapping at 30 m spatial resolution (especially for the arid and semi-arid western U.S.), providing clear field boundaries and irrigation frequency information that are vitally important for accurate agricultural water use analysis.

Self-Supervised 3D Reconstruction and Ego-Motion Estimation Via On-Board Monocular Video

Recovering the three-dimensional structure information from a monocular camera is significant for automated driving, robot navigation, and traffic safety assessment. Recent work has solved various tight issues on self-supervised monocular depth estimation of leveraging on-board videos, such as occlusion/disocclusion, dynamic objects, and scale inconsistent. Nevertheless, rare work focuses on the model’s prediction confidence and underlying relations between depths, while they are essential for a decision-making system and performance improvement, respectively. This paper proposes a novel scheme, that of correlation-aware structure, to dig into the relations between depths, converting the independent depths into a graph-like connected depth map. Subsequently, a Gaussian estimator is devised to predict the depth map and uncertainty map concurrently. The uncertainty map can show us problematic regions where it is difficult to predict, from which we further develop uncertainty-based strategies to improve the performance. Specifically, we propose a simple image preprocessing method to overcome the gradient locality issue caused by low-texture, especially in smooth roads and shadows. Also, to avoid the influence of high-uncertainty regions, we propose a solidity-aware mask to recognize the reliable pixels for training in the image. The experiments on the KITTI dataset show that our method results in a competitive performance in both depth and ego-motion estimation tasks compared with the state-of-the-art methods. Besides, additional experiments on the Make3D and Cityscapes datasets demonstrate our method’s strong generalization capability and practicality.

Field Geometry and the Spatial and Temporal Generalization of Crop Classification Algorithms—A Randomized Approach to Compare Pixel Based and Convolution Based Methods

With the ongoing trend towards deep learning in the remote sensing community, classical pixel based algorithms are often outperformed by convolution based image segmentation algorithms. This performance was mostly validated spatially, by splitting training and validation pixels for a given year. Though generalizing models temporally is potentially more difficult, it has been a recent trend to transfer models from one year to another, and therefore to validate temporally. The study argues that it is always important to check both, in order to generate models that are useful beyond the scope of the training data. It shows that convolutional neural networks have potential to generalize better than pixel based models, since they do not rely on phenological development alone, but can also consider object geometry and texture. The UNET classifier was able to achieve the highest F1 scores, averaging 0.61 in temporal validation samples, and 0.77 in spatial validation samples. The theoretical potential for overfitting geometry and just memorizing the shape of fields that are maize has been shown to be insignificant in practical applications. In conclusion, kernel based convolutions can offer a large contribution in making agricultural classification models more transferable, both to other regions and to other years.

Comparison of Support Vector Machines and Random Forests for Corine Land Cover Mapping

Land cover information is essential in European Union spatial management, particularly that of invasive species, natural habitats, urbanization, and deforestation; therefore, the need for accurate and objective data and tools is critical. For this purpose, the European Union’s flagship program, the Corine Land Cover (CLC), was created. Intensive works are currently being carried out to prepare a new version of CLC+ by 2024. The geographical, climatic, and economic diversity of the European Union raises the challenge to verify various test areas’ methods and algorithms. Based on the Corine program’s precise guidelines, Sentinel-2 and Landsat 8 satellite images were tested to assess classification accuracy and regional and spatial development in three varied areas of Catalonia, Poland, and Romania. The method is dependent on two machine learning algorithms, Random Forest (RF) and Support Vector Machine (SVM). The bias of classifications was reduced using an iterative of randomized training, test, and verification pixels. The ease of the implementation of the used algorithms makes reproducing the results possible and comparable. The results show that an SVM with a radial kernel is the best classifier, followed by RF. The high accuracy classes that can be updated and classes that should be redefined are specified. The methodology’s potential can be used by developers of CLC+ products as a guideline for algorithms, sensors, and the possibilities and difficulties of classifying different CLC classes.

MACHINE LEARNING FOR CROHN’S DISEASE PHENOTYPE MODELING USING BIOPSY IMAGES

Abstract Background Predicting Crohn’s disease (CD) phenotype development has proven challenging due to difficulties in biopsy image interpretation of histologically similar yet biologically distinct phenotypes. At initial diagnosis, mostly CD patients are classified as B1 (inflammatory behavior), they typically either retain B1 phenotype or develop more complicated B2 (stricturing), B3 (internal penetrating), or B2/B3 phenotypes (defined by Montreal Classification). Prediction of phenotype development based on baseline biopsies can radically improve our clinical care by altering disease management. Biopsy-based image analysis via Convolutional Neural Networks (CNNs) has been successful in cancer detection, but investigation into its utility for CD phenotypes is lacking. We applied a machine learning CNN model to classify CD phenotypes and histologically normal ileal controls. Methods Baseline hematoxylin &amp; eosin (H&amp;E) stained ileal biopsy slides were obtained from the Cincinnati Children’s Hospital Medical Center’s RISK validation sub cohort. At University of Virginia, biopsy slides were digitized, and a ResNet101 CNN model was trained. High resolution images were patched into 1000x1000 pixels with a 50% overlap and then resized to 256x256 pixels for training (80-20 split was kept between training and testing sets to ensure same patient patches were not mixed). Gradient Weighted Activating Mappings (GradCAMs) were used to visualize the model’s decision making process. Results We initially trained the model for CD vs. controls where it achieved 97% accuracy in detecting controls. We further trained it for classifying CD phenotypes (n=16 B1, n=16 B2, n=4 B3, n=13 B2/B3; phenotype decision at 5 year). It displayed a higher accuracy in detecting B2 (85%) while there were overlaps in the detection of other phenotypes (Figure 1). For B2, Grad-CAM heatmaps highlighted central pink areas within the lamina propria as the model’s regions of interests which were present when other phenotypes were misclassified as B2 (Figure 2). Conclusions: Here we highlight the potential utility of a machine learning image analysis model for describing CD phenotypes using H&amp;E stained biopsies. Previous studies have shown B2 to be associated with increased activation for extracellular matrix genes (connective tissue component). Our GradCAM results support this finding as the pink central areas utilized by the model for classifying B2 could be connective tissue. Further confirmation via molecular phenotyping including Sirius Red immunohistochemistry is underway. Our work supports prediction of CD phenotypes using baseline biopsies at diagnosis and has potential to influence individualized care for children with CD.

Prediction of Active Microwave Backscatter Over Snow-Covered Terrain Across Western Colorado Using a Land Surface Model and Support Vector Machine Regression.

The main objective of this article is to develop a physically constrained support vector machine (SVM) to predict C-band backscatter over snow-covered terrain as a function of geophysical inputs that reasonably represent the relevant characteristics of the snowpack. Sentinel-1 observations, in conjunction with geophysical variables from the Noah-MP land surface model, were used as training targets and input datasets, respectively. Robustness of the SVM prediction was analyzed in terms of training targets, training windows, and physical constraints related to snow liquid water content. The results showed that a combination of ascending and descending overpasses yielded the highest coverage of prediction (15.2%) while root mean square error (RMSE) ranged from 2.06 to 2.54 dB and unbiased RMSE ranged from 1.54 to 2.08 dB, but that the combined overpasses were degraded compared with ascending-only and descending-only training target sets due to the mixture of distinctive microwave signals during different times of the day (i.e., 6 A.M. versus 6 P.M. local time). Elongation of the training window length also increased the spatial coverage of prediction (given the sparsity of the training sets), but resulted in introducing more random errors. Finally, delineation of dry versus wet snow pixels for SVM training resulted in improving the accuracy of predicted backscatter relative to training on a mixture of dry and wet snow conditions. The overall results suggest that the prediction accuracy of the SVM was strongly linked with the first-order physics of the electromagnetic response of different snow conditions.

Superpixel-Guided Discriminative Low-Rank Representation of Hyperspectral Images for Classification.

In this paper, we propose a novel classification scheme for the remotely sensed hyperspectral image (HSI), namely SP-DLRR, by comprehensively exploring its unique characteristics, including the local spatial information and low-rankness. SP-DLRR is mainly composed of two modules, i.e., the classification-guided superpixel segmentation and the discriminative low-rank representation, which are iteratively conducted. Specifically, by utilizing the local spatial information and incorporating the predictions from a typical classifier, the first module segments pixels of an input HSI (or its restoration generated by the second module) into superpixels. According to the resulting superpixels, the pixels of the input HSI are then grouped into clusters and fed into our novel discriminative low-rank representation model with an effective numerical solution. Such a model is capable of increasing the intra-class similarity by suppressing the spectral variations locally while promoting the inter-class discriminability globally, leading to a restored HSI with more discriminative pixels. Experimental results on three benchmark datasets demonstrate the significant superiority of SP-DLRR over state-of-the-art methods, especially for the case with an extremely limited number of training pixels.

Random Neighbor Pixel-Block-Based Deep Recurrent Learning for Polarimetric SAR Image Classification

Polarimetric synthetic aperture radar (PolSAR) image classification is an important part of SAR data interpretation and provides more intuitive and detailed SAR polarization information. To bridge the PolSAR data and applications, it is necessary to design a comprehensive PolSAR classification framework to achieve satisfactory results. The deep neural network (DNN) appears to be a solution for the classification issue, in which it outperforms the classical supervised classifiers under the condition of sufficient training data. However, the volume of training data will greatly limit the effectiveness of practical applications. In this article, we try to solve the dependence issue on training data in three different ways: recurrent learning, data augmentation, and postprocessing. First, the long short-term memory (LSTM) network is introduced to achieve pixel sequence learning by taking into account the spatial and polarimetric features. Second, the random neighbor pixel-block (RNPB) method is proposed to increase the number of training samples for sequence learning. Third, the conditional random field (CRF) model is employed to further improve the classification accuracy. In the experiments, three sets of PolSAR data are used to evaluate the small sample performance of the proposed classification method. With only 0.5% labeled pixels for training, the proposed RNPB-LSTM-CRF method can approach 99% overall classification accuracy for all the data sets. Compared with the existing methods, the proposed method can achieve state-of-the-art results for PolSAR image classification under the condition of 1% training samples.

Graph regularization-based joint nonnegative representation for the classification of hyperspectral images

ABSTRACT The hyperspectral image classification methods based on sparse representation and collaborative representation have achieved some satisfactory results. However, some of the representation coefficients obtained from them may be negative, which is physically difficult to explain. In addition, existing methods do not consider the local geometric relationships between pixels. To address these problems, this paper proposes a joint nonnegative representation model based on the graph regularization for the hyperspectral image classification. First, for a testing pixel, the similarity between the neighbouring pixels and the testing pixel is used to collect some neighbouring pixels in a window centred at a testing pixel and construct a new neighbourhood pixel set. Each pixel in this neighbourhood pixel set is approximated by a nonnegative linear combination of training pixels from a given over-completed training pixel set. The nonnegative constraint here can ensure to some extent that the representation coefficients of those homogeneous training pixels of the testing pixel from the same class are nonnegative, while ones of those heterogeneous training pixels from different classes are as close to zero as possible. Thus, the obtained coefficients are sparse and physically explicable. Secondly, in order to preserve the manifold structure of training pixels, we also introduce a graph-based regularization constraint. Compared with some other existing methods, our model can obtain a sparse and more discriminative representation coefficient. Furthermore, it can also reveal the local structure of training pixels. The alternative iteration optimization algorithm is devised to solve the proposed model and gives its closed-form solutions in each iteration. Experiment results on four remote sensing image data sets show the superiority of the proposed method.

Hierarchical classification of Sentinel 2-a images for land use and land cover mapping and its use for the CORINE system

The aim of this study is to investigate the potential of the Sentinel-2 satellite for land use and land cover (LULC) mapping. The commonly known supervised classification algorithms, support vector machines (SVMs), random forest (RF), and maximum likelihood (ML) classification are adopted for investigation along with a proposed hierarchical classification model based on a coordination of information on the environment land cover system. The main classes for land cover and mapping in the proposed hierarchical classification are selected as water, vegetation, built-up, and bare land in the first level, which is followed by inland water, marine water, forest/meadow, vegetated agricultural land, barren land, and nonvegetated agricultural land in the second level. The study areas for the experiments are selected as the two biggest cities of Turkey, namely Ankara and Izmir, providing a sufficient number of classes for comparison purposes. During the utilized hierarchical methodology, water and vegetation are first extracted using the normalized difference water and vegetation indices. This is followed by the selection of training pixels from the remaining classes to perform and compare different supervised learning algorithms for the first- and second-level classification in terms of accuracy rates. The experimental results first reveal that while the SVMs have close accuracy performances to those with RF, they are significantly superior to the ML classification, with an average of 8% accuracy rates for LULC mapping. Second, the hierarchical classification also gives higher performances with respect to the nonhierarchical classification, with the provided gains between 4% and 10% for class-based accuracies. The overall accuracy rates of the proposed hierarchical methodology are 85% and 84% for the first-level classes and 83% and 72% for the second-level classes, respectively, for Izmir and Ankara.

A two-stage method for spectral–spatial classification of hyperspectral images

We propose a novel two-stage method for the classification of hyperspectral images. Pixel-wise classifiers, such as the classical support vector machine (SVM), consider spectral information only. As spatial information is not utilized, the classification results are not optimal and the classified image may appear noisy. Many existing methods, such as morphological profiles, superpixel segmentation, and composite kernels, exploit the spatial information. In this paper, we propose a two-stage approach inspired by image denoising and segmentation to incorporate the spatial information. In the first stage, SVMs are used to estimate the class probability for each pixel. In the second stage, a convex variant of the Mumford–Shah model is applied to each probability map to denoise and segment the image into different classes. Our proposed method effectively utilizes both spectral and spatial information of the data sets and is fast as only convex minimization is needed in addition to the SVMs. Experimental results on three widely utilized real hyperspectral data sets indicate that our method is very competitive in accuracy, timing, and the number of parameters when compared with current state-of-the-art methods, especially when the inter-class spectra are similar or the percentage of training pixels is reasonably high.

Training Data Selection for Annual Land Cover Classification for the Land Change Monitoring, Assessment, and Projection (LCMAP) Initiative

The U.S. Geological Survey’s Land Change Monitoring, Assessment, and Projection (LCMAP) initiative involves detecting changes in land cover, use, and condition with the goal of producing land change information to improve the understanding of the Earth system and provide insights on the impacts of land surface change on society. The change detection method ingests all available high-quality data from the Landsat archive in a time series approach to identify the timing and location of land surface change. Annual thematic land cover maps are then produced by classifying time series models. In this paper, we describe the optimization of the classification method used to derive the thematic land cover product. We investigated the influences of auxiliary data, sample size, and training from different sources such as the U.S. Geological Survey’s Land Cover Trends project and National Land Cover Database (NLCD 2001 and NLCD 2011). The results were evaluated and validated based on independent data from the training dataset. We found that refining the auxiliary data effectively reduced artifacts in the thematic land cover map that are related to data availability. We improved the classification accuracy and stability considerably by using a total of 20 million training pixels with a minimum of 600,000 and a maximum of 8 million training pixels per class within geographic windows consisting of nine Analysis Ready Data tiles (450 by 450 km2). Comparisons revealed that the NLCD 2001 training data delivered the best classification accuracy. Compared to the original LCMAP classification strategy used for early evaluation (e.g., Trends training data, 20,000 samples), the optimized classification strategy improved the annual land cover map accuracy by an average of 10%.

Improving Augmented Human Intelligence to Distinguish Burkitt Lymphoma From Diffuse Large B-Cell Lymphoma Cases.

To assess and improve the assistive role of a deep, densely connected convolutional neural network (CNN) to hematopathologists in differentiating histologic images of Burkitt lymphoma (BL) from diffuse large B-cell lymphoma (DLBCL). A total of 10,818 images from BL (n = 34) and DLBCL (n = 36) cases were used to either train or apply different CNNs. Networks differed by number of training images and pixels of images, absence of color, pixel and staining augmentation, and depth of the network, among other parameters. Cases classified correctly were 17 of 18 (94%), nine with 100% of images correct by the best performing network showing a receiver operating characteristic curve analysis area under the curve 0.92 for both DLBCL and BL. The best performing CNN used all available training images, two random subcrops per image of 448 × 448 pixels, random H&E staining image augmentation, random horizontal flipping of images, random alteration of contrast, reduction on validation error plateau of 15 epochs, block size of six, batch size of 32, and depth of 22. Other networks and decreasing training images had poorer performance. CNNs are promising augmented human intelligence tools for differentiating a subset of BL and DLBCL cases.

Using High Resolution Optical Imagery to Detect Earthquake-Induced Liquefaction: The 2011 Christchurch Earthquake

Using automated supervised methods with satellite and aerial imageries for liquefaction mapping is a promising step in providing detailed and region-scale maps of liquefaction extent immediately after an earthquake. The accuracy of these methods depends on the quantity and quality of training samples and the number of available spectral bands. Digitizing a large number of high-quality training samples from an event may not be feasible in the desired timeframe for rapid response as the training pixels for each class should be typical and accurately represent the spectral diversity of that specific class. To perform automated classification for liquefaction detection, we need to understand how to build the optimal and accurate training dataset. Using multispectral optical imagery from the 22 February, 2011 Christchurch earthquake, we investigate the effects of quantity of high-quality training pixel samples as well as the number of spectral bands on the performance of a pixel-based parametric supervised maximum likelihood classifier for liquefaction detection. We find that the liquefaction surface effects are bimodal in terms of spectral signature and therefore, should be classified as either wet liquefaction or dry liquefaction. This is due to the difference in water content between these two modes. Using 5-fold cross-validation method, we evaluate performance of the classifier on datasets with different pixel sizes of 50, 100, 500, 2000, and 4000. Also, the effect of adding spectral information was investigated by adding once only the near infrared (NIR) band to the visible red, green, and blue (RGB) bands and the other time using all available 8 spectral bands of the World-View 2 satellite imagery. We find that the classifier has high accuracies (75%–95%) when using the 2000 pixels-size dataset that includes the RGB+NIR spectral bands and therefore, increasing to 4000 pixels-size dataset and/or eight spectral bands may not be worth the required time and cost. We also investigate accuracies of the classifier when using aerial imagery with same number of training pixels and either RGB or RGB+NIR bands and find that the classifier accuracies are higher when using satellite imagery with same number of training pixels and spectral information. The classifier identifies dry liquefaction with higher user accuracy than wet liquefaction across all evaluated scenarios. To improve classification performance for wet liquefaction detection, we also investigate adding geospatial information of building footprints to improve classification performance. We find that using a building footprint mask to remove them from the classification process, increases wet liquefaction user accuracy by roughly 10%.