Preprint

MaskIt: Masking for efficient utilization of incomplete public datasets for training deep learning models

Authors:
Preprints and early-stage research may not have been peer reviewed yet.
To read the file of this research, you can request a copy directly from the author.

Abstract

A major challenge in training deep learning models is the lack of high quality and complete datasets. In the paper, we present a masking approach for training deep learning models from a publicly available but incomplete dataset. For example, city of Hamburg, Germany maintains a list of trees along the roads, but this dataset does not contain any information about trees in private homes and parks. To train a deep learning model on such a dataset, we mask the street trees and aerial images with the road network. Road network used for creating the mask is downloaded from OpenStreetMap, and it marks the area where the training data is available. The mask is passed to the model as one of the inputs and it also coats the output. Our model learns to successfully predict trees only in the masked region with 78.4% accuracy.

No file available

Request Full-text Paper PDF

To read the file of this research,
you can request a copy directly from the author.

ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Global declines in insects have sparked wide interest among scientists, politicians, and the general public. Loss of insect diversity and abundance is expected to provoke cascading effects on food webs and to jeopardize ecosystem services. Our understanding of the extent and underlying causes of this decline is based on the abundance of single species or taxonomic groups only, rather than changes in insect biomass which is more relevant for ecological functioning. Here, we used a standardized protocol to measure total insect biomass using Malaise traps, deployed over 27 years in 63 nature protection areas in Germany (96 unique location-year combinations) to infer on the status and trend of local entomofauna. Our analysis estimates a seasonal decline of 76%, and mid-summer decline of 82% in flying insect biomass over the 27 years of study. We show that this decline is apparent regardless of habitat type, while changes in weather, land use, and habitat characteristics cannot explain this overall decline. This yet unrecognized loss of insect biomass must be taken into account in evaluating declines in abundance of species depending on insects as a food source, and ecosystem functioning in the European landscape.
Article
Full-text available
Urban scholars have studied street networks in various ways, but there are data availability and consistency limitations to the current urban planning/street network analysis literature. To address these challenges, this article presents OSMnx, a new tool to make the collection of data and creation and analysis of street networks simple, consistent, automatable and sound from the perspectives of graph theory, transportation, and urban design. OSMnx contributes five significant capabilities for researchers and practitioners: first, the automated downloading of political boundaries and building footprints; second, the tailored and automated downloading and constructing of street network data from OpenStreetMap; third, the algorithmic correction of network topology; fourth, the ability to save street networks to disk as shapefiles, GraphML, or SVG files; and fifth, the ability to analyze street networks, including calculating routes, projecting and visualizing networks, and calculating metric and topological measures. These measures include those common in urban design and transportation studies, as well as advanced measures of the structure and topology of the network. Finally, this article presents a simple case study using OSMnx to construct and analyze street networks in Portland, Oregon.
Article
Full-text available
To learn the forest dynamics and evaluate the ecosystem services of forest effectively, a timely acquisition of spatial and quantitative information of forestland is very necessary. Here, a new method was proposed for mapping forest cover changes by combining multi-scale satellite remote-sensing imagery with time series data. Using time series Normalized Difference Vegetation Index products derived from the Moderate Resolution Imaging Spectroradiometer images (MODIS-NDVI) and Landsat Thematic Mapper/Enhanced Thematic Mapper Plus (TM/ETM+) images as data source, a hierarchy stepwise analysis from coarse scale to fine scale was developed for detecting the forest change area. At the coarse scale, MODIS-NDVI data with 1-km resolution were used to detect the changes in land cover types and a land cover change map was constructed using NDVI values at vegetation growing seasons. At the fine scale, based on the results at the coarse scale, Landsat TM/ETM+ data with 30-m resolution were used to precisely detect the forest change location and forest change trend by analyzing time series forest vegetation indices (IFZ). The method was tested using the data for Hubei Province, China. The MODIS-NDVI data from 2001 to 2012 were used to detect the land cover changes, and the overall accuracy was 94.02 % at the coarse scale. At the fine scale, the available TM/ETM+ images at vegetation growing seasons between 2001 and 2012 were used to locate and verify forest changes in the Three Gorges Reservoir Area, and the overall accuracy was 94.53 %. The accuracy of the two layer hierarchical monitoring results indicated that the multi-scale monitoring method is feasible and reliable.
Article
Full-text available
Unqualified, the statement that approximately 1.3% of the approximately 10,000 presently known bird species have become extinct since A.D. 1500 yields an estimate of approximately 26 extinctions per million species per year (or 26 E/MSY). This is higher than the benchmark rate of approximately 1 E/MSY before human impacts, but is a serious underestimate. First, Polynesian expansion across the Pacific also exterminated many species well before European explorations. Second, three factors increase the rate: (i) The number of known extinctions before 1800 is increasing as taxonomists describe new species from skeletal remains. (ii) One should calculate extinction rates over the years since taxonomists described the species. Most bird species were described only after 1850. (iii) Some species are probably extinct; there is reluctance to declare them so prematurely. Thus corrected, recent extinction rates are approximately 100 E/MSY. In the last decades, the rate is <50 E/MSY, but would be 150 E/MSY were it not for conservation efforts. Increasing numbers of extinctions are on continents, whereas previously most were on islands. We predict a 21st century rate of approximately 1,000 E/MSY. Extinction threatens 12% of bird species; another 12% have small geographical ranges and live where human actions rapidly destroy their habitats. If present forest losses continue, extinction rates will reach 1,500 E/MSY by the century's end. Invasive species, expanding human technologies, and global change will harm additional species. Birds are poor models for predicting extinction rates for other taxa. Human actions threaten higher fractions of other well known taxa than they do birds. Moreover, people take special efforts to protect birds.
Article
The study of cities needs to become more than the sum of its parts. An international Expert Panel investigates why, and how.
Conference Paper
Each corner of the inhabited world is imaged from multiple viewpoints with increasing frequency. Online map services like Google Maps or Here Maps provide direct access to huge amounts of densely sampled, georeferenced images from street view and aerial perspective. There is an opportunity to design computer vision systems that will help us search, catalog and monitor public infrastructure, buildings and artifacts. We explore the architecture and feasibility of such a system. The main technical challenge is combining test time information from multiple views of each geographic location (e.g., aerial and street views). We implement two modules: det2geo, which detects the set of locations of objects belonging to a given category, and geo2cat, which computes the fine-grained category of the object at a given location. We introduce a solution that adapts state-of-the-art CNN-based object detectors and classifiers. We test our method on “Pasadena Urban Trees”, a new dataset of 80,000 trees with geographic and species annotations, and show that combining multiple views significantly improves both tree detection and tree species classification, rivaling human performance.
Conference Paper
There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net .
Has the Earth's sixth mass extinction already arrived?
  • Anthony D Barnosky
  • Nicholas Matzke
  • Susumu Tomiya
  • O U Guinevere
  • Brian Wogan
  • Tiago B Swartz
  • Charles Quental
  • Jenny L Marshall
  • Emily L Mcguire
  • Kaitlin C Lindsey
  • Maguire
Anthony D. Barnosky, Nicholas Matzke, Susumu Tomiya, Guinevere OU Wogan, Brian Swartz, Tiago B. Quental, Charles Marshall, Jenny L. McGuire, Emily L. Lindsey, and Kaitlin C. Maguire. 2011. Has the Earth's sixth mass extinction already arrived? Nature 471, 7336: 51-57.
Biological annihilation via the ongoing sixth mass extinction signaled by vertebrate population losses and declines
  • Gerardo Ceballos
  • Paul R Ehrlich
  • Rodolfo Dirzo
Gerardo Ceballos, Paul R. Ehrlich, and Rodolfo Dirzo. 2017. Biological annihilation via the ongoing sixth mass extinction signaled by vertebrate population losses and declines. Proceedings of the national academy of sciences 114, 30: E6089-E6096.
Rasterio: geospatial raster I/O for Python programmers
  • Sean Gillies
  • B Ward
  • A S Petersen
Sean Gillies, B. Ward, and A. S. Petersen. 2013. Rasterio: geospatial raster I/O for Python programmers. URL https://github. com/mapbox/rasterio.
Accurate segmentation of dental panoramic radiographs with U-NETS
  • Mathis Thorbjørn Louring Koch
  • Christian Perslev
  • Sami Sebastian Igel
  • Brandt
Thorbjørn Louring Koch, Mathis Perslev, Christian Igel, and Sami Sebastian Brandt. 2019. Accurate segmentation of dental panoramic radiographs with U-NETS. In 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), 15-19.