Recent research has demonstrated the potential of deep neural networks (DNNs) to accurately predict wildfire spread on a given day based upon high-dimensional explanatory data from a single preceding day, or from a time series of T preceding days. For the first time, we investigate a large number of existing data-driven wildfire modeling strategies under controlled conditions, revealing the best modeling strategies and resulting in models that achieve state-of-the-art (SOTA) accuracy for both single-day and multi-day input scenarios, as evaluated on a large public benchmark for next-day wildfire spread, termed the WildfireSpreadTS (WSTS) benchmark. Consistent with prior work, we found that models using time-series input obtained the best overall accuracy, suggesting this is an important future area of research. Furthermore, we create a new benchmark, WSTS+, by incorporating four additional years of historical wildfire data into the WSTS benchmark. Our benchmark doubles the number of unique years of historical data, expands its geographic scope, and, to our knowledge, represents the largest public benchmark for time-series-based wildfire spread prediction. Code, model weights, and data can be found at this repo.

Recent object detection models for infrared (IR) imagery are based upon deep neural networks (DNNs) and require large amounts of labeled training imagery. However, publicly available datasets that can be used for such training are limited in their size and diversity. To address this problem, we explore cross-modal style transfer (CMST) to leverage large and diverse color imagery datasets so that they can be used to train DNN-based IR image-based object detectors. We evaluate six contemporary stylization methods on four publicly-available IR datasets - the first comparison of its kind - and find that CMST is highly effective for DNN-based detectors. Surprisingly, we find that existing data-driven methods are outperformed by a simple grayscale stylization (an average of the color channels). Our analysis reveals that existing data-driven methods are either too simplistic or introduce significant artifacts into the imagery. To overcome these limitations, we propose meta-learning style transfer (MLST), which learns a stylization by composing and tuning well-behaved analytic functions. We find that MLST leads to more complex stylizations without introducing significant image artifacts and achieves the best overall detector performance on our benchmark datasets.

Recently, self-supervised learning methods have shown remarkable performance rivaling supervised approaches, particularly in the realm of computer vision. This paper addresses a gap in current literature by focusing on the application of the SwAV (Swapping Assignments between Views) model to pre-train on an extensive dataset comprising one million unlabeled multi-spectral images from Sentinel-2 and Sentinel-1 (including SAR images), we investigate the impact of SSL techniques on multispectral and SAR data as compared to RGB data using crop delineation and land cover classification downstream tasks. Our results demonstrate superior performance exhibited by the 12-channel SwAV pre-trained model compared to RGB-only encodings, underscoring the benefits of SSL for enhancing computer vision for remote sensing applications. Additionally, our findings showcase the potential of SSL for smaller datasets and downstream applications.

Recently, the first foundation model developed specifically for image segmentation tasks was developed, termed the "Segment Anything Model" (SAM). SAM can segment objects in input imagery based on cheap input prompts, such as one (or more) points, a bounding box, or a mask. The authors examined the zero-shot image segmentation accuracy of SAM on a large number of vision benchmark tasks and found that SAM usually achieved recognition accuracy similar to, or sometimes exceeding, vision models that had been trained on the target tasks. The impressive generalization of SAM for segmentation has major implications for vision researchers working on natural imagery. In this work, we examine whether SAM’s performance extends to overhead imagery problems and help guide the community’s response to its development. We examine SAM’s performance on a set of diverse and widely studied benchmark tasks. We find that SAM does often generalize well to overhead imagery, although it fails in some cases due to the unique characteristics of overhead imagery and its common target objects. We report on these unique systematic failure cases for remote sensing imagery that may comprise useful future research for the community.

We propose and show the efficacy of a new method to address generic inverse problems. Inverse modeling is the task whereby one seeks to determine the hidden parameters of a natural system that produce a given set of observed measurements. Recent work has shown impressive results using deep learning, but we note that there is a trade-off between model performance and computational time. For some applications, the computational time at inference for the best performing inverse modeling method may be overly prohibitive to its use. In seeking a faster, high-performing model, we present a new method that leverages multiple manifolds as a mixture of backward (e.g., inverse) models in a forward-backward model architecture. These multiple backwards models all share a common forward model, and their training is mitigated by generating training examples from the forward model. The proposed method thus has two innovations: 1) the multiple Manifold Mixture Network (MMN) architecture, and 2) the training procedure involving augmenting backward model training data using the forward model. We demonstrate the advantages of our method by comparing to several baselines on four benchmark inverse problems, and we furthermore provide analysis to motivate its design.

There is evidence that transformers offer state-of-the-art recognition performance on tasks involving overhead imagery (e.g., satellite imagery). However, it is difficult to make unbiased empirical comparisons between competing deep learning models, making it unclear whether, and to what extent, transformer-based models are beneficial. In this paper we systematically compare the impact of adding transformer structures into state-of-the-art segmentation models for overhead imagery. Each model is given a similar budget of free parameters, and their hyperparameters are optimized using Bayesian Optimization with a fixed quantity of data and computation time. We conduct our experiments with a large and diverse dataset comprising two large public benchmarks: Inria and DeepGlobe. We perform additional ablation studies to explore the impact of specific transformer-based modeling choices. Our results suggest that transformers provide consistent, but modest, performance improvements. We only observe this advantage however in hybrid models that combine convolutional and transformer-based structures, while fully transformer-based models achieve relatively poor performance.

Automatic object detection in overhead imagery is greatly increasing the pace at which we learn about anthropic activity across diverse fields such as economics, environmental management, and engineering. Properly-trained object detection models save significant amounts of human labor when it comes to finding objects, especially rare objects, in overhead imagery. However, applying such techniques to find rare objects typically requires a large amount of labeled imagery data (typically requiring expensive manual labeling). We generate synthetic imagery to reduce the amount of manually labeled imagery required to train models, particularly for data-constrained applications. This approach takes real, unlabeled overhead imagery and inserts artificial 3D models of objects onto the imagery. To evaluate this technique, we collected a baseline dataset of overhead imagery with wind turbines that have been manually labeled and overhead imagery that does not contain wind turbines. We then add synthetic imagery to some of the unlabeled data to create a synthetic dataset. Our results indicate that adding synthetic imagery in training achieving higher levels of recall for similar levels of precision, outperforming the baseline of only real imagery.

Thermal infrared (IR) imaging has increasingly been used for remote sensing applications, which has required the adaptation of processing techniques for “natural images” (e.g., RGB images) to this unique domain in order to accommodate such differences as texture and resolution. While Convolutional Neural Networks (CNNs) have recently been shown to perform well for classification of IR images, we consider the common scenario in which the target object is partially occluded. Recent work demonstrates that deep CNNs struggle to generalize under occlusion, and Compositional CNNs (CompNets) have been proposed as deep models to mitigate this shortcoming. In this work, we apply CompNets to IR imagery, discuss the considerations in moving from natural images to IR, and analyze performance on a dataset that has been artificially occluded. Our results indicate that CompNets do, indeed, bolster robustness to occlusion in the IR domain.

Access to electricity positively correlates with many beneficial socioeconomic outcomes in the developing world including improvements in education, health, and poverty. Efficient planning for electricity access requires information on the location of existing electric transmission and distribution infrastructure; however, the data on existing infrastructure is often unavailable or expensive. We propose a deep learning based method to automatically detect electric transmission infrastructure from aerial imagery and quantify those results with traditional object detection performance metrics. In addition, we explore two challenges to applying these techniques at scale: (1) how models trained on particular geographies generalize to other locations and (2) how the spatial resolution of imagery impacts infrastructure detection accuracy. Our approach results in object detection performance with an F1 score of 0.53 (0.47 precision and 0.60 recall). Using training data that includes more diverse geographies improves performance across the 4 geographies that we examined. Image resolution significantly impacts object detection performance and decreases precipitously as the image resolution decreases.

Recently, Convolutional Neural Networks (CNNs) have demonstrated impressive performance on several visual recognition benchmark datasets utilizing overhead imagery. However, most of these analyses performed on benchmark datasets involve testing pre-trained CNNs on imagery that was collected over roughly the same locations as the training imagery. In this work we propose a benchmark problem - termed cross-geographical domain (xGD) adaptation - designed to evaluate the performance of CNNs in which they are tested on imagery collected over previously unseen geo-locations - a more challenging and practical scenario that we term cross-domain testing. We focus this work on building segmentation due to the availability of appropriate datasets. The results indicate that CNNs generalize poorly to data processed from geographic locations that were not present in training. Surprisingly, we found that larger models (pre-trained on ImageNet) generalize as well as small models in cross-domain testing, and sometimes better. This work provides the first comprehensive results for cross-domain recognition, raising awareness of this important problem. We hope that xGD can serve as a benchmark for future work; xGD uses publicly-available data, and we release our design details with this publication.

Convolutional Neural Networks (CNNs) have dominated performance on benchmark problems for object recognition in remote sensing imagery. However, recent work has shown that they may perform poorly when tested on imagery collected over a geographic location that was not present in its training imagery. In this work we explore the potential of designing synthetic overhead imagery to match a target geographic region (e.g., city or county), after which the synthetic imagery could be used to train deep learning models to recognize the unique visual features of objects/background in the target geographic location. We term this approach geographic domain matching. Towards this goal, we utilize a publicly-available dataset of synthetic overhead imagery, Synthinel-1. We systematically alter individual visual features of Synthinel-1 in an effort to match one real-world testing city: Vienna, Austria. We then evaluate whether these individual alterations improve the performance benefits of Synthinel-1 on Vienna and other cities. The results suggest that our proposed methodologies for altering the synthetic imagery to match Vienna were effective, thereby taking first step towards developing methods for designing synthetic overhead imagery for domain matching.

We consider the task of solving generic inverse problems, where one wishes to determine the hidden parameters of a natural system that will give rise to a particular set of measurements. Recently many new approaches based upon deep learning have arisen, generating promising results. We conceptualize these models as different schemes for efficiently, but randomly, exploring the space of possible inverse solutions. As a result, the accuracy of each approach should be evaluated as a function of time rather than a single estimated solution, as is often done now. Using this metric, we compare several state-of-the-art inverse modeling approaches on four benchmark tasks: two existing tasks, a new 2-dimensional sinusoid task, and a challenging modern task of meta-material design. Finally, inspired by our conception of the inverse problem, we explore a simple solution that uses a deep neural network as a surrogate (i.e., approximation) for the forward model, and then uses backpropagation with respect to the model input to search for good inverse solutions. Variations of this approach - which we term the neural adjoint (NA) - have been explored recently on specific problems, and here we evaluate it comprehensively on our benchmark. We find that the addition of a simple novel loss term - which we term the boundary loss - dramatically improves the NA’s performance, and it consequentially achieves the best (or nearly best) performance in all of our benchmark scenarios.

Recently deep learning - namely convolutional neural networks (CNNs) - have yielded impressive performance for the task of building segmentation on large overhead (e.g., satellite) imagery benchmarks. However, these benchmark datasets only capture a small fraction of the variability present in real-world overhead imagery, limiting the ability to properly train, or evaluate, models for real-world application. Unfortunately, developing a dataset that captures even a small fraction of real-world variability is typically infeasible due to the cost of imagery, and manual pixel-wise labeling of the imagery. In this work we develop an approach to rapidly and cheaply generate large and diverse synthetic overhead imagery for training segmentation CNNs. Using this approach, we generate and publicly-release a collection of synthetic overhead imagery, termed Synthinel-1, with full pixel-wise building labels. We use several benchmark datasets to demonstrate that Synthinel-1 is consistently beneficial when used to augment real-world training imagery, especially when CNNs are tested on novel geographic locations or conditions.

Segmentation convolutional neural networks (CNNs) are now popular for the semantic segmentation (i.e., dense pixel-wise labeling) of remote sensing imagery, such as color or hyperspectral satellite imagery. In recent years a large number of hand-labeled datasets of overhead imagery have emerged, leading to breakthrough performance for CNNs. However, these datasets are typically used in isolation of one another because they are either (i) annotated with heterogeneous object type labels, or (ii) they are collected over different geographic areas. This imposes a major bottleneck on the value of these datasets. In this work we present what we call a class-asymmetric loss function that makes it possible to train a single multi-class network using multiple datasets that are heterogeneously-labeled. We show, for example, that it is possible to train a segmentation algorithm for Buildings, roads, and background using two datasets: one annotated with buildings and one annotated with buildings. We propose a class asymmetric loss that under certain common conditions, allows for one to train models on datasets in which the target class is unlabeled.

Segmentation convolutional neural networks (SCNNs) are now popular for the semantic segmentation (i.e., dense pixel-wise labeling) of remote sensing imagery, such as color or hyperspectral satellite imagery. One desirable property of SCNNs when applied to remote sensing problems is rotational equivariance. This property implies that the class label assigned to a particular pixel (building, road, etc.) does not change if the input imagery is rotated by an arbitrary angle. We argue that recently proposed methods to make rotational equivariant SCNNs fall into two broad categories: easily employed methods that are somewhat ineffective, and highly effective methods that are complicated and potentially incompatible with state-of-the-art SCNN techniques. We propose a simple addition to the standard SCNN loss function that encourages the SCNN to be rotationally equivariant, and is easily added to modern SCNNs. We test the method on the Inria building labeling dataset and compare it to the popular simple approach of adding random rotational augmentations of the input imagery during training. We show that the proposed approach (i) achieves improved equivariance and (ii) yields performance improvements on average.

In this paper, we consider the development of algorithms for the automatic detection of buried threats using ground penetrating radar (GPR) measurements. GPR is one of the most studied and successful modalities for automatic buried threat detection (BTD), and a large variety of BTD algorithms have been proposed for it. Despite this, large-scale comparisons of GPR-based BTD algorithms are rare in the literature. In this paper, we report the results of a multi-institutional effort to develop advanced BTD algorithms for a real-world GPR BTD system. The effort involved five institutions with substantial experience with the development of GPR-based BTD algorithms. In this paper, we report the technical details of the advanced algorithms submitted by each institution, representing their latest technical advances, and many state-of-the-art GPR-based BTD algorithms. We also report the results of evaluating the algorithms from each institution on the large experimental data set used for development. The experimental data set comprised 120 000 m2of GPR data using surface area, from 13 different lanes across two U.S. test sites. The data were collected using a vehicle-mounted GPR system, the variants of which have supplied data for numerous publications. Using these results, we identify the most successful and common processing strategies among the submitted algorithms, and make recommendations for GPR-based BTD algorithm design.

In this work, we investigate strategies for training convolutional neural networks (CNNs) to perform recognition on remote sensing imagery. In particular we consider the particular problem of semantic segmentation in which the goal is to obtain a dense pixel-wise labeling of the input imagery. Remote sensing imagery is usually stored in the form of very large images, called “tiles”, which are too big to be segmented directly using most CNNs and their associated hardware. Therefore smaller sub-images, called “patches”, must be extracted from the available tiles. A popular strategy in the literature is to randomly sample patches from the tiles. However, in this work we demonstrate experimentally that extracting patches randomly from a uniform, non-overlapping spatial grid, leads to more accurate models. Our findings suggest the performance improvements are the result of reducing redundancy within the training dataset. We also find that sampling mini-batches of patches (for stochastic gradient descent) using constraints that maximizes the diversity of images within each batch leads to more accurate models. For example, in this work we constrained patches to come from varying tiles, or cities. These simple strategies contributed to our winning entry (in terms of overall performance) in the first year of the INRIA Building Labeling Challenge.

In this work we consider the application of convolutional neural networks (CNNs) for the semantic segmentation of remote sensing imagery (e.g., aerial color or hyperspectral imagery). In segmentation the goal is to provide a dense pixel-wise labeling of the input imagery. However, remote sensing imagery is usually stored in the form of very large images, called “tiles”, which are too large to be segmented directly using most CNNs and their associated hardware. During label inference (i.e., obtaining labels for a new large tile) smaller sub-images, called “patches”, are extracted uniformly over a tile and the resulting label maps are “stitched” (or concatenated) to create a tile-sized label map. This approach suffers from computational inefficiency and risks of discontinuities at the boundaries between the output of individual patches. In this work we propose a simple alternative approach in which the input size of the CNN is dramatically increased only during label inference. We evaluate the performance of the proposed approach against a standard stitching approach using two popular segmentation CNN models on the INRIA building labeling dataset. The results suggest that the proposed approach substantially reduces label inference time, while also yielding modest overall label accuracy increases. This approach also contributed to our winning entry (overall performance) in the INRIA building labeling competition.

Over the recent years, there has been an increasing interest in large-scale classification of remote sensing images. In this context, the Inria Aerial Image Labeling Benchmark has been released online in December 2016. In this paper, we discuss the outcomes of the first year of the benchmark contest, which consisted in dense labeling of aerial images into building / not building classes, covering areas of five cities not present in the training set. We present four methods with the highest numerical accuracies, all four being convolutional neural network approaches. It is remarkable that three of these methods use the U-net architecture, which has thus proven to become a new standard in image dense labeling.

This paper presents a methodology for automatically estimating the energy consumption of buildings from aerial imagery using data from Gainesville, Florida. By detecting buildings in the imagery using convolutional neural networks and extracting features from those building annotations, we use only imagery-derived features to estimate building energy consumption using random forests regression. For individual buildings, we achieve a predictive R2value of 0.26, and with spatial aggregation over an area of 400m×400m our predictive R2value increases to 0.95. We also explore the sensitivity of these estimates to errors in the building estimation process. Our results indicate that information limited to the size and shape of buildings, provides substantial predictive potential for the energy consumption of buildings.

Recently, convolutional neural networks (CNNs) have received substantial attention in the literature for object recognition (e.g., buildings and roads) in several remote sensing data modalities (e.g., aerial color imagery). Although CNNs have exhibited excellent recognition performance, recent research suggests that trained CNNs can often perform very poorly when applied to data collected over new geographic regions, and for which little labeled training data is available. In this work, we consider the adversarial discriminative domain adaptation (ADDA) approach to address this limitation, due its recent success on related problems. A limitation of ADDA is that it is unsupervised, so in this work we extend ADDA to a semi-supervised algorithm, in which we assume that both labeled and unlabeled data are available in the new domain (e.g., in new geographic region to be evaluated). We compare semi-supervised ADDA to ADDA and a standard fine-tuning approach wherein available labeled data is used for standard CNN training. We perform experiments on two remote sensing datasets and the results indicate that semi-supervised ADDA consistently improves over the other approaches when small amounts of labeled training data are available in the new domain.

Forward-looking ground-penetrating radar (FLGPR) has recently been investigated as a remote-sensing modality for buried target detection (e.g., landmines). In this context, raw FLGPR data are beamformed into images, and then, computerized algorithms are applied to automatically detect subsurface buried targets. Most existing algorithms are supervised, meaning that they are trained to discriminate between labeled target and nontarget imagery, usually based on features extracted from the imagery. A large number of features have been proposed for this purpose; however, thus far it is unclear as to which are the most effective. The first goal of this paper is to provide a comprehensive comparison of detection performance using existing features on a large collection of FLGPR data. Fusion of the decisions resulting from processing each feature is also considered. The second goal of this paper is to investigate two modern feature learning approaches from the object recognition literature: the bag-of-visual words and the Fisher vector for FLGPR processing. The results indicate that the new feature learning approaches lead to the best performing FLGPR algorithm. The results also show that fusion between existing features and new features yields no additional performance improvements.

Ground-penetrating radar (GPR) is one of the most popular and successful sensing modalities that have been investigated for landmine and subsurface threat detection. Many of the detection algorithms applied to this task are supervised and therefore require labeled examples of threat and nonthreat data for training. Training data most often consist of 2-D images (or patches) of GPR data, from which features are extracted and provided to the classifier during training and testing. Identifying desirable training and testing locations to extract patches, which we term “keypoints,” is well established in the literature. In contrast, however, a large variety of strategies have been proposed regarding keypoint utilization (e.g., how many of the identified keypoints should be used at threat, or nonthreat, locations). Given a variety of keypoint utilization strategies that are available, it is very unclear: 1) which strategies are best or 2) whether the choice of strategy has a large impact on classifier performance. We address these questions by presenting a taxonomy of existing utilization strategies and then evaluating their effectiveness on a large data set using many different classifiers and features. We analyze the results and propose a new strategy, called PatchSelect, which outperforms other strategies across all experiments.

In this work we consider the problem of developing algorithms that automatically identify small-scale solar photovoltaic arrays in high resolution aerial imagery. Such algorithms potentially offer a faster and cheaper solution to collecting small-scale photovoltaic (PV) information, such as their location, capacity, and the energy they produce. Here we build on previous algorithmic work by employing convolutional neural networks (CNNs), which have recently yielded major improvements in other image object recognition problems. We propose a CNN architecture for our recognition problem and then measure its detection performance on the same (publicly available) dataset that was used in previous publications. The results indicate that the CNN yields substantial performance improvements over previous results. We also investigate the recently popular approach of pre-training for CNNs.

In this work, we consider the problem of detecting target objects in remote sensing imagery; such as detecting rooftops, trees, or cars in color/hyperspectral imagery. Many detection algorithms for this problem work by assigning a decision statistic (or “confidence”) to all, or a subset, of spatial locations in the data. A threshold is then applied to the statistics to identify detections. The detection theory underpinning this general approach assumes that a yes/no decision must be made, individually, for each location. In some applications, however, the precise location of the detected objects may be less important than knowing how many total objects there are. In this work we propose two methods that can permit a generic detection algorithm to gradually lower the spatial certainty, or resolution, of its detections in order to improve the accuracy of the overall number of detected objects. We validate the proposed methods on a controlled synthetic dataset as well as a real dataset from previously published work on solar photovoltaic array detection in color aerial imagery.

In this work, the problem of developing algorithms that automatically infer information about small-scale solar photovoltaic (PV) arrays in high resolution aerial imagery is considered. Such algorithms potentially offer a faster and cheaper solution to collecting small-scale PV information, such as their location and capacity. Existing work on this topic has focused on the automatic identification and annotation of panels in the aerial imagery. We extend this work by showing that we can reliably infer the capacity of PV arrays given only (i) color aerial imagery and (ii) a precise annotation of the array location. First we demonstrate that accurate capacity estimates can be obtained simply by estimating the visible surface area of a solar array, regardless of tilt. We then build a more sophisticated model where we use additional image information related to the properties of the solar array to further improve the capacity predictions. We use a dataset of 362 manually annotated Google Earth images of solar arrays with known electricity generation capacity in North Carolina to measure the predictive performance of our models.

Most state-of-the-art dictionary learning algorithms (DLAs) are iterative, and must begin with an initial estimate of the dictionary, referred to as the seed. A seed can be generated randomly, but it has been shown that choosing a more intelligent seed often yields a better solution. For example, a seed inferred using data from a related problem, or one handcrafted based on a priori knowledge of the problem at hand can yield better solutions. Seed dictionaries appear to encode valuable a priori information however, most DLAs discard the seed after initialization. This work investigates the questions of whether the information encoded in a good seed can be leveraged further, by potentially using the seed to influence learning after initialization. This is achieved by modifying the popular DLA K-SVD to use the seed as a prior during learning, by penalizing differences between the learned dictionary and the seed. The resulting algorithm, referred to as Seed Shrinkage Dictionary Learning (SSDL), is examined against K-SVD on image denoising experiments using several benchmark images. The results indicate that utilizing the seed as a prior in this way consistently yields improved denoising performance in our experiments. This simple approach motivates the development of more sophisticated approaches that leverage a priori information in useful seeds.

The ground penetrating radar (GPR) is a popular and successful remote sensing modality that has been investigated for landmine detection. GPR offers excellent detection performance, but it is limited by a low rate of advance (ROA) due to its short sensing standoff distance. Standoff distance refers to the distance between the sensing platform and the location in front of the platform where the GPR senses the ground. Large standoff (high ROA) sensing modalities have been investigated as alternatives to the GPR, but they do not yet achieve comparable detection performance. This paper proposes a new sensor management approach, called multistate management (MSM), which combines large and short standoff sensors on the same platform in a way that leverages their respective advantages, yielding a system with better ROA and detection performance. MSM is more difficult to analyze than traditional systems because it allows sensor activity and system velocity to change over time. Therefore, a new probabilistic model based on queuing theory, called Q-MSM, is also proposed for analyzing and designing detection systems operating with MSM. Simulations were conducted using real field-collected data for a system with a large standoff forward-looking infrared camera and a GPR. The system is operated with MSM, and the results show that this leads to better ROA and detection performance than can be attained otherwise. Furthermore, the results show that Q-MSM can accurately predict the behavior of the MSM system, validating its utility for analyzing and designing such systems.

Optimal management of anemia due to End-Stage Renal Disease (ESRD) is a challenging task to physicians due to large inter-subject variability in response to Erythropoiesis Stimulating Agents (ESA). We demonstrate that an optimal dosing strategy for ESA can be derived using Reinforcement Learning (RL) techniques. In this study, we show some preliminary results of using a batch RL method, called Fitted Q-Iteration, to derive optimal ESA dosing strategies from retrospective treatment data. Presented results show that such dosing strategies are superior to a standard ESA protocol employed by our dialysis facilities.

In this paper, the effect of class imbalance in the case base of a case-based classifier is investigated as it pertains to case base reduction and the resulting classifier performance. A k-nearest neighbor algorithm is used as a classifier and the random mutation hill climbing (RMHC) algorithm is used for case base reduction. The effects at various levels of positive class prevalence are tested in a binary classification problem. The results indicate that class imbalance is detrimental to both case base reduction and classifier performance. Selection with RMHC generally improves the classification performance regardless of the case base prevalence.