Uncertainty Map Research Articles

Insights Into the Aerodynamic Versus Radiometric Surface Temperature Debate in Thermal‐Based Evaporation Modeling

AbstractGlobal evaporation monitoring from Earth observation thermal infrared satellite missions is historically challenged due to the unavailability of any direct measurements of aerodynamic temperature. State‐of‐the‐art one‐source evaporation models use remotely sensed radiometric surface temperature as a substitute for the aerodynamic temperature and apply empirical corrections to accommodate for their inequality. This introduces substantial uncertainty in operational drought mapping over complex landscapes. By employing a non‐parametric model, we show that evaporation can be directly retrieved from thermal satellite data without the need of any empirical correction. Independent evaluation of evaporation in a broad spectrum of biome and aridity yielded statistically significant results when compared with eddy covariance observations. While our simplified model provides a new perspective to advance spatio‐temporal evaporation mapping from any thermal remote sensing mission, the direct retrieval of aerodynamic temperature also generates the highly required insight on the critical role of biophysical interactions in global evaporation research.

A novel multichannel deep learning model for fast denoising of Monte Carlo dose calculations: preclinical applications

Objective. In preclinical radiotherapy with kilovolt (kV) x-ray beams, accurate treatment planning is needed to improve the translation potential to clinical trials. Monte Carlo based radiation transport simulations are the gold standard to calculate the absorbed dose distribution in external beam radiotherapy. However, these simulations are notorious for their long computation time, causing a bottleneck in the workflow. Previous studies have used deep learning models to speed up these simulations for clinical megavolt (MV) beams. For kV beams, dose distributions are more affected by tissue type than for MV beams, leading to steep dose gradients. This study aims to speed up preclinical kV dose simulations by proposing a novel deep learning pipeline. Approach. A deep learning model is proposed that denoises low precision (∼106 simulated particles) dose distributions to produce high precision (109 simulated particles) dose distributions. To effectively denoise the steep dose gradients in preclinical kV dose distributions, the model uses the novel approach to use the low precision Monte Carlo dose calculation as well as the Monte Carlo uncertainty (MCU) map and the mass density map as additional input channels. The model was trained on a large synthetic dataset and tested on a real dataset with a different data distribution. To keep model inference time to a minimum, a novel method for inference optimization was developed as well. Main results. The proposed model provides dose distributions which achieve a median gamma pass rate (3%/0.3 mm) of 98% with a lower bound of 95% when compared to the high precision Monte Carlo dose distributions from the test set, which represents a different dataset distribution than the training set. Using the proposed model together with the novel inference optimization method, the total computation time was reduced from approximately 45 min to less than six seconds on average. Significance. This study presents the first model that can denoise preclinical kV instead of clinical MV Monte Carlo dose distributions. This was achieved by using the MCU and mass density maps as additional model inputs. Additionally, this study shows that training such a model on a synthetic dataset is not only a viable option, but even increases the generalization of the model compared to training on real data due to the sheer size and variety of the synthetic dataset. The application of this model will enable speeding up treatment plan optimization in the preclinical workflow.

Bayesian Gabor Network With Uncertainty Estimation for Pedestrian Lane Detection in Assistive Navigation

Automatic pedestrian lane detection is a challenging problem that is of great interest in assistive navigation and autonomous driving. Such a detection system must cope well with variations in lane surfaces and illumination conditions so that a vision-impaired user can navigate safely in unknown environments. This paper proposes a new lightweight Bayesian Gabor Network (BGN) for camera-based detection of pedestrian lanes in unstructured scenes. In our approach, each Gabor parameter is represented as a learnable Gaussian distribution using variational Bayesian inference. For the safety of vision-impaired users, in addition to an output segmentation map, the network provides two full-resolution maps of aleatoric uncertainty and epistemic uncertainty as well-calibrated confidence measures. Our Gabor-based method has fewer weights than the standard CNNs, therefore it is less prone to overfitting and requires fewer operations to compute. Compared to the state-of-the-art semantic segmentation methods, the BGN maintains a competitive segmentation performance while achieving a significantly compact model size (from <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.8\times $ </tex-math></inline-formula> to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$237.6\times $ </tex-math></inline-formula> reduction), a fast prediction time (from <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$1.2\times $ </tex-math></inline-formula> to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$67.5\times $ </tex-math></inline-formula> faster), and a well-calibrated uncertainty measure. We also introduce a new lane dataset of 10,000 images for objective evaluation in pedestrian lane detection research.

Fully automated segmentation of clinical target volume in cervical cancer from magnetic resonance imaging with convolutional neural network.

PurposeContouring clinical target volume (CTV) from medical images is an essential step for radiotherapy (RT) planning. Magnetic resonance imaging (MRI) is used as a standard imaging modality for CTV segmentation in cervical cancer due to its superior soft‐tissue contrast. However, the delineation of CTV is challenging as CTV contains microscopic extensions that are not clearly visible even in MR images, resulting in significant contour variability among radiation oncologists depending on their knowledge and experience. In this study, we propose a fully automated deep learning–based method to segment CTV from MR images.MethodsOur method begins with the bladder segmentation, from which the CTV position is estimated in the axial view. The superior–inferior CTV span is then detected using an Attention U‐Net. A CTV‐specific region of interest (ROI) is determined, and three‐dimensional (3‐D) blocks are extracted from the ROI volume. Finally, a CTV segmentation map is computed using a 3‐D U‐Net from the extracted 3‐D blocks.ResultsWe developed and evaluated our method using 213 MRI scans obtained from 125 patients (183 for training, 30 for test). Our method achieved (mean ± SD) Dice similarity coefficient of 0.85 ± 0.03 and the 95th percentile Hausdorff distance of 3.70 ± 0.35 mm on test cases, outperforming other state‐of‐the‐art methods significantly (p‐value < 0.05). Our method also produces an uncertainty map along with the CTV segmentation by employing the Monte Carlo dropout technique to draw physician's attention to the regions with high uncertainty, where careful review and manual correction may be needed.ConclusionsExperimental results show that the developed method is accurate, fast, and reproducible for contouring CTV from MRI, demonstrating its potential to assist radiation oncologists in alleviating the burden of tedious contouring for RT planning in cervical cancer.

Geostatistical Analysis of Mangrove Ecosystem Health: Mapping and Modelling of Sampling Uncertainty Using Kriging

This study assessed the health of the mangrove ecosystem and mapped the spatial variation in selected variables sampled across the Matang Mangrove Forest Reserve (MMFR) by using a geostatistical technique. A total of 556 samples were collected from 56 sampling points representing mangrove biotic and abiotic variables. All variables were used to generate the semivariogram model. The predicted variables over the entire MMFR have an overall prediction accuracy of 85.16% (AGB), 90.78% (crab abundance), 97.3% (soil C), 99.91% (soil N), 89.23% (number of phytoplankton species), 95.62% (number of diatom species), 99.36% (DO), and 87.33% (turbidity). Via linear weight combination, the prediction map shows that mangrove ecosystem health in Kuala Trong throughout the Sungai Kerang is excellent (5: MQI &gt; 1.5). Some landward areas of Kuala Trong were predicted to have moderate health (3: −0.5 ≤ MQI ≤ 0.5), while Kuala Sepetang was predicted to have the bad ecosystem health (2: −1.5 ≤ MQI ≤ −0.5), with active timber harvesting operations and anthropogenic activities in the landward areas. The results of this method can be utilised to carry out the preferred restoration, through appropriate management and facilities distribution, for improving the ecosystem health of mangroves.

Forest tree species distribution for Europe 2000-2020: mapping potential and realized distributions using spatiotemporal machine learning.

This article describes a data-driven framework based on spatiotemporal machine learning to produce distribution maps for 16 tree species (Abies alba Mill., Castanea sativa Mill., Corylus avellana L., Fagus sylvatica L., Olea europaea L., Picea abies L. H. Karst., Pinus halepensis Mill., Pinus nigra J. F. Arnold, Pinus pinea L., Pinus sylvestris L., Prunus avium L., Quercus cerris L., Quercus ilex L., Quercus robur L., Quercus suber L. and Salix caprea L.) at high spatial resolution (30 m). Tree occurrence data for a total of three million of points was used to train different algorithms: random forest, gradient-boosted trees, generalized linear models, k-nearest neighbors, CART and an artificial neural network. A stack of 305 coarse and high resolution covariates representing spectral reflectance, different biophysical conditions and biotic competition was used as predictors for realized distributions, while potential distribution was modelled with environmental predictors only. Logloss and computing time were used to select the three best algorithms to tune and train an ensemble model based on stacking with a logistic regressor as a meta-learner. An ensemble model was trained for each species: probability and model uncertainty maps of realized distribution were produced for each species using a time window of 4 years for a total of six distribution maps per species, while for potential distributions only one map per species was produced. Results of spatial cross validation show that the ensemble model consistently outperformed or performed as good as the best individual model in both potential and realized distribution tasks, with potential distribution models achieving higher predictive performances (TSS = 0.898, R2logloss = 0.857) than realized distribution ones on average (TSS = 0.874, R2logloss = 0.839). Ensemble models for Q. suber achieved the best performances in both potential (TSS = 0.968, R2logloss = 0.952) and realized (TSS = 0.959, R2logloss = 0.949) distribution, while P. sylvestris (TSS = 0.731, 0.785, R2logloss = 0.585, 0.670, respectively, for potential and realized distribution) and P. nigra (TSS = 0.658, 0.686, R2logloss = 0.623, 0.664) achieved the worst. Importance of predictor variables differed across species and models, with the green band for summer and the Normalized Difference Vegetation Index (NDVI) for fall for realized distribution and the diffuse irradiation and precipitation of the driest quarter (BIO17) being the most frequent and important for potential distribution. On average, fine-resolution models outperformed coarse resolution models (250 m) for realized distribution (TSS = +6.5%, R2logloss = +7.5%). The framework shows how combining continuous and consistent Earth Observation time series data with state of the art machine learning can be used to derive dynamic distribution maps. The produced predictions can be used to quantify temporal trends of potential forest degradation and species composition change.

Assessing Machine Learning-Based Prediction under Different Agricultural Practices for Digital Mapping of Soil Organic Carbon and Available Phosphorus

Predicting soil chemical properties such as soil organic carbon (SOC) and available phosphorus (Ava-P) content is critical in areas where different land uses exist. The distribution of SOC and Ava-P is influenced by both natural and anthropogenic factors. This study aimed at (1) predicting SOC and Ava-P in a piedmont plain of Northeast Iran using the Random Forests (RF) and Cubist mathematical models and hybrid models (Regression Kriging), (2) comparing the models’ results, and (3) identifying the key variables that influence the spatial dynamics of soil SOC and Ava-P under different agricultural practices. The machine learning models were trained with 201 composite surface soil samples and 24 ancillary data, including climate (C), organism (O), topography- relief (R), parent material (P) and key soil features (S) according to the SCORPAN digital soil mapping framework, which can predictively represent soil formation factors spatially. Clay, one of the most critical soil properties with a well-known relationship to SOC, was the most important predictor of SOC, followed by open-access multispectral satellite images-based vegetation and soil indices. Ava-P had a similar set of effective variables. Hybrid approaches did not improve model accuracy significantly, but they did reduce map uncertainty. In the validation set, Ava-P was calculated using the RF algorithm with a normalized root mean square (NRMSE) of 96.8, while SOC was calculated using the Cubist algorithm with an NRMSE of 94.2. These values did not change when using the hybrid technique for Ava-P; however, they changed just by 1% for SOC. The management of SOC content and the supply of Ava-P in agricultural activities can be guided by SOC and Ava-P digital distribution maps. Produced digital maps in which the soil scientist plays an active role can be used to identify areas where concentrations are high and need to be protected, where uncertainty is high and sampling is required for further monitoring.

A new digital elevation model (DEM) dataset of the entire Antarctic continent derived from ICESat-2

Abstract. Antarctic digital elevation models (DEMs) are essential for fieldwork, ice motion tracking and the numerical modelling of the ice sheet. In the past 30 years, several Antarctic DEMs derived from satellite data have been published. However, these DEMs either have coarse spatial resolution or aggregate observations spanning several years, which limit their further scientific applications. In this study, the new generation satellite laser altimeter Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) is used to generate a new Antarctic DEM for both the ice sheet and ice shelves. Approximately 4.69 × 109 ICESat-2 measurement points from November 2018 to November 2019 are used to estimate surface elevations at resolutions of 500 m and 1 km based on a spatiotemporal fitting method. Approximately 74 % of Antarctica is observed and the remaining observation gaps are interpolated using the normal kriging method. The DEM is formed from the estimated elevations in 500 m and 1 km grid cells, and is finally posted at the resolution of 500 m. National Aeronautics and Space Administration (NASA) Operation IceBridge (OIB) airborne data are used to evaluate the generated Antarctic DEM (hereafter called the ICESat-2 DEM) in individual Antarctic regions and surface types. Overall, a median bias of −0.19 m and a root-mean-square deviation of 10.83 m result from approximately 5.2 × 106 OIB measurement points. The accuracy and uncertainty of the ICESat-2 DEM vary in relation to the surface slope and roughness, and more reliable estimates are found in the flat ice sheet interior. The ICESat-2 DEM is comparable to other DEMs derived from altimetry, stereophotogrammetry and interferometry. Similar results are found when comparing to elevation measurements from kinematic Global Navigation Satellite System (GNSS) (GPS and the Russian GLONASS) transects. The elevations of high accuracy and ability of annual updates make the ICESat-2 DEM an addition to the existing Antarctic DEM groups, and it can be further used for other scientific applications. The generated ICESat-2 DEM (including the map of uncertainty) can be downloaded from National Tibetan Plateau Data Center, Institute of Tibetan Plateau Research, Chinese Academy of Sciences at https://data.tpdc.ac.cn/en/disallow/9427069c-117e-4ff8-96e0-4b18eb7782cb/ (last access: 27 June 2022) (Shen et al., 2021a, https://doi.org/10.11888/Geogra.tpdc.271448).

Hand-feel soil texture observations to evaluate the accuracy of digital soil maps for local prediction of soil particle size distribution: A case study in Central France

Digital maps of soil properties are now widely available. End-users now can access several digital soil mapping (DSM) products of soil properties, produced using different models, calibration/training data, and covariates at various spatial scales from global to local. Therefore, there is an urgent need to provide easy-to-understand tools to communicate map uncertainty and help end-users assess the reliability of DSM products for use at local scales. In this study, we used a large amount of hand-feel soil texture (HFST) data to assess the performance of various published DSM products on the prediction of soil particle size distribution in Central France. We tested four DSM products for soil texture prediction developed at various scales (global, continental, national, and regional) by comparing their predictions with approximately 3 200 HFST observations realized on a 1:50 000 soil survey conducted after release of these DSM products. We used both visual comparisons and quantitative indicators to match the DSM predictions and HFST observations. The comparison between the low-cost HFST observations and DSM predictions clearly showed the applicability of various DSM products, with the prediction accuracy increasing from global to regional predictions. This simple evaluation can determine which products can be used at the local scale and if more accurate DSM products are required.

Robust Depth Completion with Uncertainty-Driven Loss Functions

Recovering a dense depth image from sparse LiDAR scans is a challenging task. Despite the popularity of color-guided methods for sparse-to-dense depth completion, they treated pixels equally during optimization, ignoring the uneven distribution characteristics in the sparse depth map and the accumulated outliers in the synthesized ground truth. In this work, we introduce uncertainty-driven loss functions to improve the robustness of depth completion and handle the uncertainty in depth completion. Specifically, we propose an explicit uncertainty formulation for robust depth completion with Jeffrey's prior. A parametric uncertain-driven loss is introduced and translated to new loss functions that are robust to noisy or missing data. Meanwhile, we propose a multiscale joint prediction model that can simultaneously predict depth and uncertainty maps. The estimated uncertainty map is also used to perform adaptive prediction on the pixels with high uncertainty, leading to a residual map for refining the completion results. Our method has been tested on KITTI Depth Completion Benchmark and achieved the state-of-the-art robustness performance in terms of MAE, IMAE, and IRMSE metrics.

About Tree Height Measurement: Theoretical and Practical Issues for Uncertainty Quantification and Mapping

Forest height is a fundamental parameter in forestry. Tree height is widely used to assess a site’s productivity both in forest ecology research and forest management. Thus, a precise height measure represents a necessary step for the estimation of carbon storage at the local, national, and global scales. In this context, error in height measurement necessarily affects the accuracy of related estimates. Ordinarily, forest height is surveyed by ground sampling adopting hypsometers. The latter suffers from many errors mainly related to the correct tree apex identification (not always well visible in dense stands) and to the measurement process itself. In this work, a statistically based operative method for estimating height measurement uncertainty (σH) was proposed using the variance propagation law. Some simulations were performed involving several combinations of terrain slope, tree height, and survey distances by modelling the σH behaviour and its sensitivity to such parameters. Results proved that σH could vary between 0.5 m and up to 20 m (worst case). Sensitivity analysis shows that terrain slopes and distance poorly affect σH, while angles are the main drivers of height uncertainty. Finally, to give a practical example of such deductions, tree height uncertainty was mapped at the global scale using Google Earth Engine and summarized per forest biomes. Results proved that tropical biomes have higher uncertainty (from 1 m to 4 m) while shrublands and tundra have the lowest (under 1 m).

Direct quantification of epistemic and aleatoric uncertainty in 3D U-net segmentation.

Purpose: To derive a multinomial probability function and quantitative measures of the data and epistemic uncertainty as direct output of a 3D U-Net segmentation network. Approach: A set of T1 brain MRI images were downloaded from the Connectome Project and segmented using FMRIB's FAST algorithm to be used as ground truth. A 3D U-Net neural network was trained with sample sizes of 200, 500, and 898 T1 brain images using a loss function defined as the negative logarithm of the likelihood based on a derivation of the definition of the multinomial probability function. From this definition, the epistemic and aleatoric uncertainty equations were derived and used to quantify maps of the uncertainty along with tissue segmentations. Results: Maps of the tissue segmentation along with the epistemic and aleatoric uncertainty, per voxel, are presented. The uncertainty decreased based on the increasing number of training data used to train the neural network. The neural network trained with 898 volumes resulted in uncertainty maps that were high primarily in the tissue boundary regions. The epistemic and aleatoric uncertainty were averaged over all test data (connectome and tumor separately), and the epistemic uncertainty showed a decreasing trend, as expected, with increasing numbers of data used to train the model. The aleatoric uncertainty showed a similar trend which was also expected as the aleatoric uncertainty is not expected to be as dependent on the number of training data. Conclusion: The derived uncertainty equations from a multinomial probability distribution were able to quantify the aleatoric and epistemic uncertainty per voxel and are applicable for all two-dimensional and three-dimensional neural networks.

Understanding Uncertainty Maps in Vision with Statistical Testing.

Quantitative descriptions of confidence intervals and uncertainties of the predictions of a model are needed in many applications in vision and machine learning. Mechanisms that enable this for deep neural network (DNN) models are slowly becoming available, and occasionally, being integrated within production systems. But the literature is sparse in terms of how to perform statistical tests with the uncertainties produced by these overparameterized models. For two models with a similar accuracy profile, is the former model's uncertainty behavior better in a statistically significant sense compared to the second model? For high resolution images, performing hypothesis tests to generate meaningful actionable information (say, at a user specified significance level ) is difficult but needed in both mission critical settings and elsewhere. In this paper, specifically for uncertainties defined on images, we show how revisiting results from Random Field theory (RFT) when paired with DNN tools (to get around computational hurdles) leads to efficient frameworks that can provide a hypothesis test capabilities, not otherwise available, for uncertainty maps from models used in many vision tasks. We show via many different experiments the viability of this framework.

Stitching interferometry of cylindrical surface

High resolution and high precision are advantages of measuring cylindrical objects employing stitching interferometry. However, due to errors in the cylinder position in a measurement procedure, it is difficult to ensure the accuracy of a macroscopic shape contour and surface waviness contour of the workpiece of the stitching results. To overcome this problem, we proposed a new strategy of stitching interferometry for cylindricity calibration. The measuring process is divided into two parts - to measure the macroscopic shape and waviness contour, respectively. Finally, the two measurement results are merged to produce an accurate and complete 360° form map. Besides, we build a two-dimensional Gaussian filter weighting function for a cylindrical profile according to its special shape characteristic. To express the stitching results more precisely, we provide the uncertainty map and power spectral density of the stitching results. An experiment is carried out by taking a ceramic plug gage as the workpiece. The results demonstrate the excellent performance of our method in analyzing the characteristic of the cylindrical profile and show good agreement with such by using cylindricity measuring instruments.

UNCERTAINTY REPRESENTATION AND QUANTIFICATION OF 3D BUILDING MODELS

Abstract. The quality of environmental perception is of great interest for localization tasks in autonomous systems. Maps, generated from the sensed information, are often used as additional spatial references in these applications. The quantification of the map uncertainties gives an insight into how reliable and complete the map is, avoiding the potential systematic deviation in pose estimation. Mapping 3D buildings in urban areas using Light detection and ranging (LiDAR) point clouds is a challenging task as it is often subject to uncertain error sources in the real world such as sensor noise and occlusions, which should be well represented in the 3D models for the downstream localization tasks. In this paper, we propose a method to model 3D building façades in complex urban scenes with uncertainty quantification, where the uncertainties of windows and façades are indicated in a probabilistic fashion. The potential locations of the missing objects (here: windows) are inferred by the available data and layout patterns with the Monte Carlo (MC) sampling approach. The proposed 3D building model and uncertainty measures are evaluated using the real-world LiDAR point clouds collected by Riegl Mobile Mapping System. The experimental results show that our uncertainty representation conveys the quality information of the estimated locations and shapes for the modelled map objects.

Perception-Aware Planning for Active SLAM in Dynamic Environments

This paper presents a perception-aware path planner for active SLAM in dynamic environments using micro-aerial vehicles (MAV). The “Next-Best-View” planner (NBVP planner) is combined with an active loop closing, which is called the Active Loop Closing Planner (ALCP planner). The planner is proposed to avoid both static and dynamic obstacles in unknown environments while reducing the uncertainty of the SLAM system and further improving the accuracy of localization. First, the receding horizon strategy is adopted to find the next waypoint. The cost function that combines the exploration gain and the loop closing gain is designed. The former can reduce the mapping uncertainty, while the latter takes the loop closing possibility into consideration. Second, a key waypoint selection strategy is designed. The selected key waypoints, instead of all waypoints, are treated as potential loop-closing points to make the algorithm more efficient. Moreover, a fuzzy RRT-based dynamic obstacle avoidance algorithm is adopted to realize obstacle avoidance in dynamic environments. Simulations in different challenging scenarios are conducted to verify the effectiveness of the proposed algorithm.

Making the US national forest inventory spatially contiguous and temporally consistent

Signatories to the United Nations Framework Convention on Climate Change (UNFCCC) are required to annually report economy-wide greenhouse gas emissions and removals, including the forest sector. National forest inventory (NFI) is considered the main source of data for reporting forest carbon stocks and changes to UNFCCC. However, NFI samples are often collected asynchronously across regions in intervals of 5–10 years or sub-sampled annually, both introducing uncertainties in estimating annual carbon stock changes by missing a wide range of forest disturbance and recovery processes. Here, we integrate satellite observations with forest inventory data across the conterminous United States to improve the spatial and temporal resolution of NFI for estimating annual carbon stocks and changes. We used more than 120 000 permanent plots from the US forest inventory and analysis (FIA) data, surveyed periodically at sampling rate of 15%–20% per year across the US to develop non-parametric remote sensing-based models of aboveground biomass carbon density (AGC) at 1 ha spatial resolution for the years 2005, 2010, 2015, 2016, and 2017. The model provided a relatively unbiased estimation of AGC compared to ground inventory estimates at plot, county, and state scales. The uncertainty of the biomass maps and their contributions to estimates of forest carbon stock changes at county and state levels were quantified. Our results suggest that adding spatial and temporal dimensions to the forest inventory data, will significantly improve the accuracy and precision of carbon stocks and changes at jurisdictional scales.

Polar transform network for prostate ultrasound segmentation with uncertainty estimation.

Automatic and accurate prostate ultrasound segmentation is a long-standing and challenging problem due to the severe noise and ambiguous/missing prostate boundaries. In this work, we propose a novel polar transform network (PTN) to handle this problem from a fundamentally new perspective, where the prostate is represented and segmented in the polar coordinate space rather than the original image grid space. This new representation gives a prostate volume, especially the most challenging apex and base sub-areas, much denser samples than the background and thus facilitate the learning of discriminative features for accurate prostate segmentation. Moreover, in the polar representation, the prostate surface can be efficiently parameterized using a 2D surface radius map with respect to a centroid coordinate, which allows the proposed PTN to obtain superior accuracy compared with its counterparts using convolutional neural networks while having significantly fewer (18%∼41%) trainable parameters. We also equip our PTN with a novel strategy of centroid perturbed test-time augmentation (CPTTA), which is designed to further improve the segmentation accuracy and quantitatively assess the model uncertainty at the same time. The uncertainty estimation function provides valuable feedback to clinicians when manual modifications or approvals are required for the segmentation, substantially improving the clinical significance of our work. We conduct a three-fold cross validation on a clinical dataset consisting of 315 transrectal ultrasound (TRUS) images to comprehensively evaluate the performance of the proposed method. The experimental results show that our proposed PTN with CPTTA outperforms the state-of-the-art methods with statistical significance on most of the metrics while exhibiting a much smaller model size. Source code of the proposed PTN is released at https://github.com/DIAL-RPI/PTN.

OC-0771 Uncertainty map for error prediction in deep learning-based head and neck tumor auto-segmentation

OC-0771 Uncertainty map for error prediction in deep learning-based head and neck tumor auto-segmentation

Pedometric Mapping of Soil Classes: A Case Study of San Mateo de Otao, Peru

Conventional soil maps are designed based on expert criteria, a characteristic that reduces their reproducibility and generates subjective uncertainty. Pedometric mapping uses mathematical and statistical principles, which makes it the opposite of conventional mapping. It was proposed to apply the pedometric mapping in San Mateo de Otao and find out its characteristics against the conventional one. Satellite and field data were used to extract covariables (soil-forming factors) and soil classes. The data were modeled with Naïve Bayes, global uncertainty was calculated by resubstitution, cross-validation and retention, and local uncertainty with the confusion and Shannon indices. A low uncertainty map was obtained with six identified soil classes, relief, and parent material having the most important covariates. We conclude that pedometric mapping has considerable advantages over conventional mapping and its application is possible under the context of soil survey in Peru.