A generalizable and accessible approach to machine learning with global satellite imagery.
Ontology highlight
ABSTRACT:
SUBMITTER: Rolf E
PROVIDER: S-EPMC8292408 | biostudies-literature | 2021 Jul
REPOSITORIES: biostudies-literature
Publications
A generalizable and accessible approach to machine learning with global satellite imagery.
Rolf Esther E Proctor Jonathan J Carleton Tamma T Bolliger Ian I Shankar Vaishaal V Ishihara Miyabi M Recht Benjamin B Hsiang Solomon S
Nature communications 20210720 1
Combining satellite imagery with machine learning (SIML) has the potential to address global challenges by remotely estimating socioeconomic and environmental conditions in data-poor regions, yet the resource requirements of SIML limit its accessibility and use. We show that a single encoding of satellite imagery can generalize across diverse prediction tasks (e.g., forest cover, house price, road length). Our method achieves accuracy competitive with deep neural networks at orders of magnitude ...[more]
Similar Datasets
Project description:Machine learning approaches using satellite imagery are providing accessible ways to infer socioeconomic measures without visiting a region. However, many algorithms require integration of ground-truth data, while regional data are scarce or even absent in many countries. Here we present our human-machine collaborative model which predicts grid-level economic development using publicly available satellite imagery and lightweight subjective ranking annotation without any ground data. We applied the model to North Korea and produced fine-grained predictions of economic development for the nation where data is not readily available. Our model suggests substantial development in the country's capital and areas with state-led development projects in recent years. We showed the broad applicability of our model by examining five of the least developed countries in Asia, covering 400,000 grids. Our method can both yield highly granular economic information on hard-to-visit and low-resource regions and can potentially guide sustainable development programs.
| S-EPMC10603027 | biostudies-literature
Project description:Recognizing the importance of road infrastructure to promote human health and economic development, actors around the globe are regularly investing in both new roads and road improvements. However, in many contexts there is a sparsity-or complete lack-of accurate information regarding existing road infrastructure, challenging the effective identification of where investments should be made. Previous literature has focused on overcoming this gap through the use of satellite imagery to detect and map roads. In this piece, we extend this literature by leveraging satellite imagery to estimate road quality and concomitant information about travel speed. We adopt a transfer learning approach in which a convolutional neural network architecture is first trained on data collected in the United States (where data is readily available), and then "fine-tuned" on an independent, smaller dataset collected from Nigeria. We test and compare eight different convolutional neural network architectures using a dataset of 53,686 images of 2,400 kilometers of roads in the United States, in which each road segment is measured as "low", "middle", or "high" quality using an open, cellphone-based measuring platform. Using satellite imagery to estimate these classes, we achieve an accuracy of 80.0%, with 99.4% of predictions falling within the actual or an adjacent class. The highest performing base model was applied to a preliminary case study in Nigeria, using a dataset of 1,000 images of paved and unpaved roads. By tailoring our US-model on the basis of this Nigeria-specific data, we were able to achieve an accuracy of 94.0% in predicting the quality of Nigerian roads. A continuous case estimate also showed the ability, on average, to predict road quality to within 0.32 on a 0 to 3 scale (with higher values indicating higher levels of quality).
| S-EPMC8270213 | biostudies-literature
Project description:The accurate and comprehensive mapping of land cover has become a central task in modern environmental research, with increasing emphasis on machine learning approaches. However, a clear technical definition of the land cover class is a prerequisite for learning and applying a machine learning model. One of the challenging classes is naturalness and human influence, yet mapping it is important due to its critical role in biodiversity conservation, habitat assessment, and climate change monitoring. We present an interpretable machine learning approach to map patterns related to territorial protected and anthropogenic areas as proxies of naturalness and human influence using satellite imagery. To achieve this, we train a weakly-supervised convolutional neural network and subsequently apply attribution methods such as Grad-CAM and occlusion sensitivity mapping. We propose a novel network architecture that consists of an image-to-image network and a shallow, task-specific head. Both sub-networks are connected by an intermediate layer that captures high-level features in full resolution, allowing for detailed analysis with a wide range of attribution methods. We further analyze how intermediate layer activations relate to their attributions across the training dataset to establish a consistent relationship. This makes attributions consistent across different scenes and allows for a large-scale analysis of remote sensing data. The results highlight that our approach is a promising way to observe and assess naturalness and territorial protection.
| S-EPMC10725256 | biostudies-literature
Project description:This study aims to develop a software framework for predicting aridity using vegetation indices (VI) from LANDSAT 8 OLI images. VI data are predicted using machine learning (ml): Random Forest (RF) and Correlation and Regression Trees (CART). Comparison of prediction using Artificial Neural Network (ANN), Support Vector Machine (SVM), k-nearest neighbors (k-nn) and Multivariate Adaptive Regression Spline (MARS). Prediction results are interpolated using Inverse Distance Weight (IDW). This study was conducted in stages: (1) Image preprocessing; (2) calculating numerical data extracted from the LANDSAT band imagery using vegetation indices; (3) analyzing correlation coefficients between VI; (4) prediction using RF and CART; (5) comparing performances between RF and CART using ANN, SVM, k-nn, and MARS; (6) testing the accuracy of prediction using Mean Square Error (MSE) and Mean Absolute Percentage Error (MAPE); (7) interpolating with IDW. Correlation coefficient of VI data shows a positive correlation, the lowest r (0.07) and the highest r (0.98). The experiments show that the RF and CART algorithms have efficiency and effectivity in determining the aridity areas better than the ANN, SVM, k-nn, and MARS algorithm. RF has a difference between the predicted results and 1.04% survey data MAPE and the smallest value close to zero is 0.05 MSE. CART has a difference between the predicted results and 1.05% survey data MAPE and the smallest value approaching to zero which is 0.05 MSE. The prediction results of VI show that in 2020 most of the study areas were low vegetation areas with the Normalized Difference Vegetation Index (NDVI) < 0.21, had an indication of drought with the Vegetation Health Index (VHI) < 31.10, had a Vegetation Condition Index (VCI) in some areas between 35%-50% (moderate drought) and < 35% (high drought). The Burn Area Index (dBAI) values are between -3, 971 and -2,376 that show the areas have a low fire risk, and index values are between -0, 208 and -0,412 that show the areas are starting vegetation growth. The result of this study shows that the machine learning algorithms is an accurate and stable algorithm in predicting the risks of drought and land fire based on the VI data extracted from the LANDSAT 8 OLL imagery. The VI data contain the record of vegetation condition and its environment, including humidity, temperatures, and the environmental vegetation health.
| S-EPMC8157165 | biostudies-literature
Project description:BackgroundEarly and accurate diagnosis of Alzheimer's disease (AD) is essential for disease management and therapeutic choices that can delay disease progression. Machine learning (ML) approaches have been extensively used in attempts to develop algorithms for reliable early diagnosis of AD, although clinical usefulness, interpretability, and generalizability of the classifiers across datasets and MRI protocols remain limited.MethodsWe report a multi-diagnostic and generalizable approach for mild cognitive impairment (MCI) and AD diagnosis using structural MRI and ML. Classifiers were trained and tested using subjects from the AD Neuroimaging Initiative (ADNI) database (n = 570) and the Open Access Series of Imaging Studies (OASIS) project database (n = 531). Several classifiers are compared and combined using voting for a decision. Additionally, we report tests of generalizability across datasets and protocols (IR-SPGR and MPRAGE), the impact of using graph theory measures on diagnostic classification performance, the relative importance of different brain regions on classification for better interpretability, and an evaluation of the potential for clinical applicability of the classifier.ResultsOur "healthy controls (HC) vs. AD" classifier trained and tested on the combination of ADNI and OASIS datasets obtained a balanced accuracy (BAC) of 90.6% and a Matthew's correlation coefficient (MCC) of 0.811. Our "HC vs. MCI vs. AD" classifier trained and tested on the ADNI dataset obtained a 62.1% BAC (33.3% being the by-chance cut-off) and 0.438 MCC. Hippocampal features were the strongest contributors to the classification decisions (approx. 25-45%), followed by temporal (approx. 13%), cingulate, and frontal regions (approx. 8-13% each), which is consistent with our current understanding of AD and its progression. Classifiers generalized well across both datasets and protocols. Finally, using graph theory measures did not improve classification performance.ConclusionsIn sum, we present a diagnostic tool for MCI and AD trained using baseline scans and a follow-up diagnosis regardless of progression, which is multi-diagnostic, generalizable across independent data sources and acquisition protocols, and with transparently reported performance. Rated as potentially clinically applicable, our tool may be clinically useful to inform diagnostic decisions in dementia, if successful in real-world prospective clinical trials.
| S-EPMC9347083 | biostudies-literature
Project description:With the advancement of global civilisation, monitoring and managing dumpsites have become essential parts of environmental governance in various countries. Dumpsite locations are difficult to obtain in a timely manner by local government agencies and environmental groups. The World Bank shows that governments need to spend massive labour and economic costs to collect illegal dumpsites to implement management. Here we show that applying novel deep convolutional networks to high-resolution satellite images can provide an effective, efficient, and low-cost method to detect dumpsites. In sampled areas of 28 cities around the world, our model detects nearly 1000 dumpsites that appeared around 2021. This approach reduces the investigation time by more than 96.8% compared with the manual method. With this novel and powerful methodology, it is now capable of analysing the relationship between dumpsites and various social attributes on a global scale, temporally and spatially.
| S-EPMC10015540 | biostudies-literature
Project description:Highly accurate nighttime cloud detection in satellite imagery is challenging due to the absence of visible to near-infrared (0.38-3 μm, VNI) data, which is critical for distinguishing clouds from other ground features. Fortunately, Machine learning (ML) techniques can more effectively leverage the limited wavelength information and show high-accuracy cloud detection based on vast sample volume. However, accurately distinguishing cloud pixels solely through thermal infrared bands (8-14 μm, TIR) is challenging, and acquiring numerous, high-quality and representative samples of nighttime images for ML proves to be unattainable in practice. Given the thermal infrared radiation transmission process and the fact that daytime and nighttime have the same source of radiance in TIR, we propose a sample generation idea that uses daytime images to provide samples of nighttime cloud detection, which is different from the traditional sample construction methods (e.g., manually label, simulation and transfer learning method), and can obtain samples effectively. Based on this idea, nighttime cloud detection experiments were carried out for MODIS, GF-5 (02) and Himawari-8 satellites, respectively. The results were validated by the Lidar cloud product and manual labels and show that our nighttime cloud detection result has higher accuracy than MYD35 (78.17%) and the ML model trained by nighttime manual labels (75.86%). The accuracy of the three sensors is 82.19%, 88.71%, and 79.34%, respectively. Moreover, we validated and discussed the performance of our algorithm on various surface types (vegetation, urban, barren and water). The results revealed that the accuracy of the three sensors over barren was found to be poor and varied with surface types but overall high. Our study can provide a novel perspective on nighttime cloud detection of muti-spectral satellite imagery.
| S-EPMC11561103 | biostudies-literature
Project description:IntroductionThis article proposes a novel method for matching places based on visual similarity, using high-resolution satellite imagery and machine learning. This approach strengthens comparisons when the built environment is a potential confounder, as in many injury research studies.MethodsAs an example, I apply this method to study the spatial influence of alcohol outlets (AOs) on firearm violence in Philadelphia, Pennsylvania, specifically beer stores and bar/restaurants. Using a case-control framework, city blocks with shootings in 2017-2018 were matched with similar-looking blocks with no shootings, based on analysis with a pretrained convolutional neural network and t-distributed stochastic neighbour embedding. Logistic regression was used to estimate the OR of a shooting on the same block as an AO and within one-block and two-block distances, conditional on additional factors such as land use, demographic composition and illegal drug activity.ResultsThe case-control matches were similar in visual appearance, on human inspection, and were well balanced on covariate measures. The fully adjusted model estimated an increased shootings risk for locations with beer stores within one block, OR=1.5, 95% CI 1.1 to 2.1, p=0.02, and locations with bar/restaurants on the same block, OR=1.6, 95% CI 1.1 to 2.4, p=0.02.ConclusionThese findings align with previous study findings while addressing the concern that AOs might systematically be located in certain kinds of environments, providing stronger evidence of a causal effect on nearby firearm violence. Matching on visual similarity can improve observational injury studies involving place-based risks.
| S-EPMC8268046 | biostudies-literature
Project description:Radar (SAR) satellites systematically acquire imagery that can be used for volcano monitoring, characterising magmatic systems and potentially forecasting eruptions on a global scale. However, exploiting the large dataset is limited by the need for manual inspection, meaning timely dissemination of information is challenging. Here we automatically process ~ 600,000 images of > 1000 volcanoes acquired by the Sentinel-1 satellite in a 5-year period (2015-2020) and use the dataset to demonstrate the applicability and limitations of machine learning for detecting deformation signals. Of the 16 volcanoes flagged most often, 5 experienced eruptions, 6 showed slow deformation, 2 had non-volcanic deformation and 3 had atmospheric artefacts. The detection threshold for the whole dataset is 5.9 cm, equivalent to a rate of 1.2 cm/year over the 5-year study period. We then use the large testing dataset to explore the effects of atmospheric conditions, land cover and signal characteristics on detectability and find that the performance of the machine learning algorithm is primarily limited by the quality of the available data, with poor coherence and slow signals being particularly challenging. The expanding dataset of systematically acquired, processed and flagged images will enable the quantitative analysis of volcanic monitoring signals on an unprecedented scale, but tailored processing will be needed for routine monitoring applications.Supplementary informationThe online version contains supplementary material available at 10.1007/s00445-022-01608-x.
| S-EPMC9633547 | biostudies-literature
Project description:The worldwide growth of cancer incidence can be explained in part by changes in the prevalence and distribution of risk factors. There are geographical gaps in the estimates of cancer prevalence, which could be filled with innovative methods. We used deep learning (DL) features extracted from satellite images to predict cancer prevalence at the census tract level in seven cities in the United States. We trained the model using detailed cancer prevalence estimates from 2018 available in the CDC (Center for Disease Control) 500 Cities project. Data from 3500 census tracts covering 14,483,366 inhabitants were included. Features were extracted from 170,210 satellite images with deep learning. This method explained up to 64.37% (median = 43.53%) of the variation of cancer prevalence. Satellite features are highly correlated with individual socioeconomic and health measures that are linked to cancer prevalence (age, smoking and drinking status, and obesity). A higher similarity between two environments is associated with better generalization of the model (p = 1.10-6). This method can be used to accurately estimate cancer prevalence at a high spatial resolution without using surveys at a fraction of the cost.
| S-EPMC7766226 | biostudies-literature