More Automated Spatial Deep Learning: The Picterra Tool
There is no question deep learning and artificial intelligence techniques have transformed remote sensing, computer vision, and spatial analysis. Until now, most efforts would have had to code their efforts, segment or semantically segment data, and then also layer and parallelize their code to run on high performance or cloud-based systems. While this may not be a major issue for those with software engineering backgrounds, it was a restriction for those interested in conducting spatial and remote sensing analysis to have these additional skills.
A new tool, called Picterra (https://picterra.ch/) which was discussed by Julien Rebetez in a recent Mapscaping podcast, enables a relatively easy to use interface that allows users to upload remote sensing images whereby users can identify and train an automated detector to find and detect objects of interest. This means that Web Map Service (WMS) and other raster data could be used directly for deep learning-based spatial analysis by those with minimal experience in artificial intelligence techniques. This has the potential to open up large and often open source datasets to the wider user community.
How Picterra Works
The basic premise of Picterra is that you can upload raster data, you then identify and train a detector to find relevant objects of interest, identify objects for measuring accuracy to compare your results to, and then apply the created model to your dateset.
The tool also provides capabilities to integrate links to web services or data repositories that enable analysis for objects of interest in remote datasets. Common data include satellite or unmanned aerial vehicle (UAV) data. Data could be flat image files or multispectral data, allowing analysis to utilize the visible light and non-visible part of the electromagnetic spectrum. Data from WMS can be directly integrated for analysis.
The key method in the application is an object detection technique that uses deep learning neural networks to train on objects users simply click and identify using drawn polygons.
Generally, users apply the application through an iterative process by selecting polygons of interest and training the tool until a desired level of accuracy and data sensitivity is achieved. The approach works well in identifying object and pixel data, enabling visual signatures and context data to be used together. Identifying complex shapes is one benefit from the deep learning technique, where multiple stages of analysis that detect edges and then more complex shapes and objects through a layered neural network is applied.
Additionally, Picterra has created a plugin interface with ArcGIS and QGIS to enable the two popular GIS platforms to be used in analysis. Although these plugins are still at a relatively early stage of development, it does mean Picterra can be used directly as part of GIS efforts.
For desired shapes or objects that are relatively simple in shape or have minimal diversity, it could be possible to have small training sets, such as under 100 objects, to produce a very accurate model that can identify most objects of interest.
The overall output is a set of polygons that can be counted or quantified for area and volume, among other measures, and exported as common shapefiles, csv, or klm files. The advantage for users it also gives them the power of a cloud-based service to parallelize the application and process with minimal setup.
A nice feature of having the tool work on a cloud-based system is that once you submit your job, then you can shutoff your work computer, even for large data problems, and by the next day you can have the answer to your question.
Benefits and the Future of Picterra
Perhaps the best benefit of Picterra is that many industries could benefit from this tool because users aren’t required to have an indepth knowledge of deep learning methods and access to cloud or high performance services.
Industries and fields such as manufacturing, urban planning, agriculture, forestry, and even real estate are among those that can benefit from a tool that can tell you, in one afternoon, answers to problems that could have taken days or even months.
The developments by Picterra are similar to what we have seen in recent years, which is efforts to make deep learning more accessible to the spatial community. Tools such as Orfeo, RoboSat, Solaris, and even ArcGIS’s Image Server are examples of open source and proprietary efforts that have also attempted to facilitate user access to computational capabilities for deep learning and computer vision spatial analysis., . Related: Platforms for Making Deep Learning Easier for GIS
Picterra’s tool has greatly improved this area by making the effort even more simple. The application of deep learning for geospatial analysis still requires subject-matter knowledge and some, potentially minimal coding to enable a successful project, but by lowering the entry point for analysts, it is clear that more tools will now seek to make the process even easier. As we see more tools coming out that incorporate deep learning as part of their methodology provided to users interested in remote sensing and spatial analysis, we can expect that these tools will try to make it even easier to conduct analysis that until recently has been restricted to relatively few experts and scientists.