remote sensing

Oil-Storage Tank Instance Segmentation

Pioneering the concept of synthetic frequency bands for remote sensing applications and beyond! SFBs highlight hidden radiometric signatures in multispectral imagery through sophisticated transforms that make targeted objects clearly distinguishable with respect to their surroundings. SFBs are computed using a linear-complexity algorithm with minimal hardware requirements and in real-time in case of 4k video. The demo shows a pansharp RGB image of a residential neighborhood in Ankara, Turkey (first), the SFB targeting buildings with red roofs (second), and the CSL semantic triplet (https://lnkd.in/dgJXtaX) developed at the European Commission's Joint Research Centre (jrc) for the detection & segmentation of building footprints (third). More to come soon.

Offshore platform detection and segmentation

Oil-tanker ship detection and segmentation

Fracking site detection and segmentation

Oil tank detection and segmentation

Road segment extraction is used along with stitching methods for road network mapping. The method presented segments the image space (aerial/drone imagery) in the absence of training data for the purpose of identifying relevant road sections.

The PANchromatic Clustering with Adaptive KErnels algorithm for unsupervised satellite image segmentation using max-tree based differential area profiles. Images are WorldView 3 © DigitalGlobe, Inc, panchromatic acquisitions at native resolution and show test areas in Johannesburg, South Africa. The PANCAKE algorithm extracts the most prominent features in a fully unsupervised and globally consistent manner (on any image, any resolution, any spectral set, any modality) converting pixels into AI-ready material.

Automatic built-up detection from multi-spectral imagery. The method employs an extended version of the original PANTEX algorithm customized for 8-band data-sets from the WorldView-2 and WorldView-3 satellites. The input images are required to be atmospherically compensated. The deliverable is a binary layer in which foreground areas (white color) coincide with built-up (built-up extent layer), and background areas (black color) with everything else. The algorithm is available as a task (protogenV2PANTEX10) on the GBDX platform of DigitalGlobe.

Unsupervised building footprint extraction using Alpha-Tree Differential Attribute Profiles. The method makes use of the latest, state of the art hierarchical image representation data-structure, the Alpha-Tree. A tree polychotomy scheme know as “attribute zoning” is computed from which Differential Attribute Profile (DAP) vector fields can be accessed. The Alpha-Trees are computed from vector data (multi-spectral imagery) thus the entries of the new DAPs contain dissimilarity values that optimize material separation and lead to accurate and automatic segmentation based on size, shape and radiometry. The method employs additional layers such as the unsupervised LULC and unsupervised built-up extent, for false positive reduction and self-supervised learning. The method concludes in < 3min for typical WorldView-2 multi-spectral data-sets of approximate coverage of 120km x 20km @ 2m spatial resolution. This does not include target contour refinement and vectorization. The image shows a blend in of the original rgb layer and the segmentation result. Object intensity (gray-scale) indicates confidence [0%-100%] that the segmented object is a building. The scene shows the south sector of the city of Kano, Nigeria.

Unsupervised building footprint extraction from DigitalGlobe’s 8-band WorldView 2 an 3 imagery using Differential Area Profile vector fields. The first image shows the city of Brest, France and the wider area of approximately 313 square km. The second shows the raw output (prior to ‘rectangularization’). Source article: http://www.mdpi.com/2220-9964/5/3/22/pdf

Automatic built-up detection using the CSL (characteristic-scale, saliency, level) semantic triplet.

The CSL is the result of the structured collapse of a differential area profile vector field computed from the max-tree data structure of a radiometrically reduced image. The example shows a section of Seattle (WA). More on https://lnkd.in/emYcWAB and https://lnkd.in/eRiAUZV

Aleppo, Syria. First image: rgb view of a city segment. Second image: the scale-level-response (SLR) or CSL model for built-up discovery constrained by the PANTEX layer (aggregated corner density map).

Denver (CO), south. Building footprints using differential area profiles from the max-tree and min-tree scale spaces.

Segmentation of urban structures using the CSL semantic layer

The CSL is a compression scheme for Differential Attribute Profile (DAP) vector fields that reduces feature spaces of hundreds or thousands of features to a set of 3; the scale, saliency and level. Fusion of these tree using a variant of the RGB2HSV algorithm enables the discrimination and unsupervised segmentation of objects as shown in the example above.

The images show a panchromatic view of Brisbane, Australia and the corresponding CSL layer in which urban structures are colored according to size, contrast and proximity to the next size-class.

The method is globally consistent and does not require any fine tuning. It is based on the max-tree algorithm and runs in sequential, shared memory, distributed memory and cloud platforms. It has been used to populate the Global Human Settlement Layer of the European Commission with building footprints from pixels originating from non-homogeneous sources.

maxtree attribute filters for remote sensing

An example of a size & shape sensitive attribute filter for search space reduction in building footprint detection using the max-tree algorithm.

Oil Tank Segmentation.

A simple method for segmenting oil tanks or other similar structures is presented using morphological connected attribute filters implemented on the Max-Tree data structure.

The specific example shows a cluster of three oil tanks in Denver, CO. The image is a WorldView-2@DigitalGlobe panchromatic acquisition at a spatial resolution of 0.5 m. The method utilizes size and compactness (99%) filter criteria, and the tree structure is configured with standard connectivity and the subtractive filtering rule.

Automated oil tank extraction using PROTOGEN V3 on the Geospatial Big Data Platform (GBDX) of DigitalGlobe.

Unsupervised satellite image segentation for building footprint extraction. The method used is the Attribute Maximization segmentation strategy for the Alpha-Tree algorithm with structural compactness selected as the maximization attribute. The tree search space is further constrained using size thresholds to mininmize the bias of small or very large connected components.

Location: Ankara, Turkey Geography: 39°59'15.0"N, 32°51'00.6"E Source imagery: DigitalGlobe

Alpha-Tree Segmentation

The exercise demonstrates the Non-Target Clustering method of the Alpha-Tree algorithm for anomaly detection. In this case, since airplanes differ substantially in shape and size between them, instead of trying to find a robust rule to segment them, we cluster the rather homogeneous background and remove clusters of odd shapes (like roads, buildings, etc.) that we are certain are not airplanes. Each anomaly or ‘hole’ in the clustered background coincides with a single plane.

Finding boats in VHR pan-sharp satellite/aerial imagery

using the alpha-tree non-target clustering algorithm. The method has been used for detecting and tracking maritime vessels with no beacon or other radio signatures. Can be utilized in salvage and rescue operations, monitoring illegal trafficking & maritime traffic management

Sydney LULC using WorldView 2,atmospherically compensated 8-band imagery

An improved algorithm for fully unsupervised Land-Use Land-Cover (LULC) generation using atmospherically compensated 8-band imagery from the WV2=< series. The method identifies natural elements only, i.e. vegetation, cloud, water, soil, shadows, etc.

Built-up is often the set difference between the sum of all classes the remaining pixels.

The method uses atmospherically compensated 8-band (optical+NIR) images from the WorldView-2 and WorldView-3 satellites. For each image pixel a similarity index is computed between its spectral signature (radiometric vector data) and entries from a custom spectral-signature library. LULC layers are generated by a hierarchical class assignment algorithm and are post-processed by connected and structural morphological operators for noise reduction. An intelligent blend-in method guarantees class overlap resolution and minor blank spot treatment.

The LULC method is based on a number of image analysis algorithms that do not require user input for class feature learning (machine learning) and classification. It is fully automated and concludes in < 2min for typical WorldView-2 multi-spectral full extent data sets, i.e. 120km x 20km @ 2m spatial resolution.

The LULC output contains the following classes: vegetation, bare soil, water, cloud, shadows (major), no-data zones & unclassified (mostly built-up). The LULC method is available on the GBDX platform of DigitalGlobe; task name protogenV2LULC.

Unsupervised, high precision cloud segmentation

from multi-spectral satellite imagery using the band fusion techniques and the max-tree algorithm.

Unsupervised land-use / land-cover layer for PROTOGEN V3 @ GBDX, DigitalGlobe. Notice the accuracy of the cloud mask (white). The example shows a region in Kilosa, Tanzania.

automated, real-time satellite image analytics

The standard delivery package of PROMETHEUS 3.0, atmospherically compensated imagery, land-use land cover, building footprints and built-up extent. Enjoy!

Automatic rubble detection from very high resolution aerial imagery: Haiti, Jan. 2010 case study. First: central Port-au-Prince with several destroyed buildings visible; second: the rubble density heat map; third: masked rubble clusters. The method employs Alpha-Tree segmentation and filtering based on size, compactness and standard deviation. An earlier system implementation based on Max-Tree and Min-Tree pairs is described in:

http://www.isprs.org/proceedings/2011/isrse-34/211104015Final00015.pdf

Automated rubble detection in Port-au-Prince, Haiti after the earthquake in January 2010. The images were produced by the method described in this paper

Unsupervised segmentation of mud-huts from rural African regions using Alpha-Trees. The method runs on WorldView 2 & 3 multi-spectral imagery after atmospheric compensation. It detects huts of minimal size 4 m^2.

Automatic solar panel extraction from WorldView-3 panchromatic imagery using connected morphological attribute filters. The method involves extensive connected attribute operators configured with mask-based second generation connectivity and non-increasing shape attributes. The operators are implemented on the Min-Tree data structure. All attribute thresholds are fixed and no adaptation is required. The workflow considers additional layers like the re-sampled land-use & land-cover and building outlines for false positive reduction.

Semi-automatic swimming pool segmentation extraction using PROTOGEN Radiometric Miner on the GBDX platform @ DigitalGlobe. The method makes use of positive examples (rectangles) of swimming pools. For each rectangle the swimming pool is localized and attributed with an object specific spectral signature. The mean and variance of the average signature from all positive examples is computed and stored as a reference.

In a pass through the image, the spectral signature of each pixel is compared against the mean pool signature and a similarity index is computed. The similarity range is form 0% to 100% (scalar value encoded in gray-scale).

A Max-Tree data structure is computed from the similarity map that is used for connected component labeling and attribute filtering. The latter produces a segmentation based on size, shape and similarity criteria.

The images above demonstrate successful swimming pool segmentation with accuracy threshold >= 90%. The resulted data-set contained >95% of all swimming pools in the entire AOI.

Detecting and measuring coastal change

Coastal change is a global phenomenon that is attributed to tides, powerful sea currents and overall climate change. It has the potential to threaten communities and local economies of coastal towns and cities. Accurate detection and measurement of coastal change can provide valuable data for scientific studies of coastal evolution, as well as guide flooding disaster preparedness and mitigation.

We developed an end-to-end GBDX workflow for coastal change detection and measurement at the native resolution of DigitalGlobe’s 8-band multispectral imagery that can execute at an unprecedented scale.

The gradients below are distance maps of change in steps of 2m/pixel.

Coastline gain due to tidal effects. Blue: water unchanged between the past and present images; red region: the loss in water with respect to the past image; gray: residual LULC classes. The animation shows past, present and difference image.

Coastline retreat due to sea currents and possible water level increase. Blue: water unchanged between the past and present images; green region: the gain in water with respect to the past image; gray: residual LULC classes. The animation shows past, present and difference image.

Ordered image captions

• The village of Maroantserta and surrounding areas in map view.

• The same area photographed in 2015. Muddy water is visible in the rivers and the respective deltas.

• The floods following the passing of cyclone Enawo. The image was acquired 4 days following the event in 03/11/2017.

• The flood map of the viewed area. Note that this contains all types of impure waters.