Watch the webinar to learn how to employ AI and turn your aerial images into precision farming insights without any GIS or data science knowledge!
- What ATLAS is about and how it applies AI algorithms to automate data analysis tasks
- Why users don't need GIS and data science knowledge to make use of AI
- Success story #1 about the time saved on localizing weed patches within large areas of beetroot
- Success story #2 about processing airborne multispectral imaging for crop health analysis
- Success story #3 about precise tree counting.
Speaker: Alexey Yankelevich, Co-Founder and Head of Software Development, SPH Engineering
Date/Time/Language: the webinar will be available in two more languages
Q: Good annotation depends on engineer qualification.
That's true. We had another webinar describing in detail how to better annotate data to get a better result. The recording is available on our YouTube channel.
Q: You have desktop app for Atlas?
ATLAS is not a normal desktop application. Atlas is a web-based application, it works from any browser. If you need an offline installation, then you can consider Atlas on premise. Nevertheless, it is not something you can take to the field.
Q: Is it possible when using Atlas to do weed detection in crop field?
Yes, for that you need to train your detector for detection of specific weeds. Here is an example
Q: Do you have pre-made detectors? For example, if I'm too lazy to train my own?
No, we do not have this option. We recommend to train your own detector, because your conditions might be unique.
Q: If I understand correctly, the software works from drone images. Is there a minimum specification for the image/type of drone used?
Atlas supports images from drones and satellites. It must be in JPEG or GeoTIFF format.
Q: When starting using Atlas, are there some videos explaining how to proceed ?
We have published detailed tutorial videos on SPH Engineering’s YouTube channel . Also, you may access video manuals on Atlas home page .
Q: Does Atlas provide an API?
Yes, we have some basic REST API, and we keep working on it. If you have a specific project in mind, contact us, and we will discuss the availability of certain services as API calls.
Q: With regards to detectors - what code base would the detectors be developed in?
For those who use Atlas through our user interface, all this stuff is hidden under the hood. You don't need to use any programming language to train your own detector. You just have to use the user interface for making annotations. We have already developed a neural network architecture which is optimal for certain tasks. All you need to do is just to feed annotations and press “Train”.
Atlas has a modular architecture. If you want to use it in on-premise installation or you want to add your own neural network to Atlas, technically this is possible through API. This is not a feature that is available right now for a cloud-based installation. So, you can build your own neural network and use Atlas as a data management system, as well as for annotation and storing results, but at the same time you can implement your own detector on any programming language which you like.
If you feel that our existing neural network architectures are not sufficient, drop us a message to discuss how we can help you with this integration.
Q: Is the Farm Management Software part of Atlas’ solutions or a third-party software?
Atlas covers territory segmentation tasks, so you can easily make crops counting, plant counting in Atlas, then export the results as vector layers to a farm management solution you prefer. The same way you will interact with a normal GIS system. There are a lot of farm management services on the market, and we do not plan to implement another one within Atlas, because there is a clear integration path between Atlas and farm management software.
Q: Can you use Atlas with DEM rasters for stockpile detection?
We can use orthomosaics for detection, and we support DEM data for volume measurements. Also this April we will introduce additional services for different volumetric tasks. Please try Atlas, and if you feel that something is missing, let us know, we are very open for such conversations.
Q: So does Atlas process RGB, thermal and Multispectral data to generate the orthomosaics, NDVI maps and other agricultural index maps?
Atlas works with RGB images. We also can stitch maps which contain near-infrared channel and stitch an orthomosaic.
Q: Is it possible to train the algorithm for different problems (i.e. weeds, malnutrition, pests, ...) simultaneously/in parallel and have them show up as separate problem areas?
Yes, you can train separate detectors for each type of problem and apply them to the same map, so that you will have separate layers for each problem.
Q: Do the JPEG images for Atlas require metadata like geospatial information?
We do not require special georeferencing metadata for JPEG images. If you do not need this georeferencing information, then you can upload JPEGs without this metadata and process these images, which even can be taken from just an ordinary camera, not a drone.
Q: Alexei is referring to train your detectors. Could you clarify what he is referring to as a detector? Is he talking about the camera on the drone or is the detector a process within the Atlas software that we must use to zoom in on a weed, for example, to train the AI to identify that weed?
Yes, the detectors are a part of Atlas software, so you upload orthomosaic and images to Atlas, and then Atlas helps you identify certain areas on the field.
Q: When building the detector, can we import shape file for the objects?
Right now you can import annotation as a GeoJSON file. Shape file will be supported shorty in one of the next Atlas releases. For now, you can simply convert shape: export to GeoJSON (almost any GIS system supports export to GeoJSON), then import to Atlas, and this object can be used as a manual annotation.
Q: What kind of Geographic Reference Systems are supported? Do I have to use only orthomosaics in WGS 84 or can i use also UTM System?
You can use almost any coordinate system. It is not required to convert orthomosaics to WGS 84 before uploading it to Atlas. As long as you upload a standardized coordinate system, most probably we can deal with it.
Q: Do you have any mapping software/hardware to control the drone during the data collection phase? If so what is the minimum type of drone used?
Yes, we do have. Our flagship product is a ground control software, which works with almost all commercially available drones on the market and helps plan flights for DJI and non-DJI drones. Please check the list of supported drones . The software allows plan proper missions in such a way, that the data will be acquired with a proper GSD and a proper overlap. Please check ugcs.com and check our YouTube channel to see how the software works.
Q: If you're on the Starter plan, can you pay for additional AI processing above 500MP, or do you need to upgrade the tier?
Right now we have fixed tears, but shortly we will introduce the ability to purchase additional packages of megapixels without changing the plan. If you feel that you have a one time job that hits the limit of your current plan, please post the message to our support team , If possible, we will sort it out even without changing your current plan.
Q: Will you need to train on a certain crop deficiency on every flight, or once per year per crop? Same with a specific weed?
Crop can change a lot during the year, and it may have a very different color and shape during the entire life cycle. Just one detector might not be sufficient for the entire lifecycle. You might need to spend one year to train several detectors for each season, but next year you will be able to use them for the same territory and the same crop. Share your workflow with our support team , and our consultants can help you better understand how to properly plan your spending on Atlas. We regularly do this type of consulting and don’t charge for it.
Q: How does the software perform when counting closely packed plants such as plantain/banana trees?
If plants are more or less of the same size, then after segmentation you can apply a postprocessing step that will separate detected areas into objects of the same size. You can find “Show as circles” option under the “Appearance configuration” button of object card.
Q: Is it possible to analyze hyperspectral images?
No, not yet, this feature is not yet in a standard package. Currently Atlas provides a tool for building detectors easily, but in case of hyperspectral data we have a lot more variations and options. If you are open for a conversation with us, we will be glad to call you and discuss hyperspectral analysis.
Q: Is Atlas able to mash satellite RGB and Multispectral images to improve spatial resolution of Multispectral images?
Q: Is it possible to share trained detector with different independent user working on Atlas?
Example. One farmer trains detector for his area, can he share his detector with different farmer having similar fields in the same are who uses Atlas with independent agreement? Of course among cloud users.
Yes, we have such a mechanism. You can train your detector and share it with your neighbor who has similar fields, for example.
Q: If we have subscribed Explorer Plus, and at one moment, we need more than 5000 MP, the subscription is switched to Business Plus or we can pay the amount of extra analyzed pixels?
We do not have yet a mechanism to obtain more pixels without changing a plan, but we plan to. If you hit this problem, let our support know that you need extra megapixels and we will sort it out. Shortly this ability will appear in the system.
Q: Does the detector take into account row or other shapes rather than color patterns of pixels during the discrimination? It would be a powerful solution for weed detection in arable crops AOB.
We do not rely just on color since this is not how deep neural networks work. We take into account many factors: colors, shapes, relative positions, different texture specifics. Also during the training process we synthesize some additional data which make the model more reliable.
Q: You showed the tree detection - what kind of resolution you recommend to detect any individual tree at what agl? In a mixed forest?
First of all, you need to understand the tree size. On a regular tree plantation in most cases you will have trees of more or less same size. For example, I demonstrated you tree counting on a satellite image, which is 50 cm per pixel, and the trees on the plantation have radius about 2 meters (or ~4 m diameter). So, if we divide 4 m by 50 cm , we have 8 pixels width. Because trees are more or less circular, we have 8 pixels width and 8 pixel height. For tree detection, you may expect from 5 to 8 pixels per such an object. This is how you can calculate he required GSD. In a mixed forest, the ability to perform tree detection will depend on distances between the trees and variations in sizes, and might be a complicated task.
In UgCS flight planning software we have a photogrammetry tool which helps plan photogrammetry missions, and one of the input parameter for these photogrammetry missions is GSD. You can roughly calculate the required GSD, following the same procedure I just demonstrated. Then you input this GSD into photogrammetry planner, and the planner in UgCS calculates the required altitude to have the desired GSD. The procedure is almost completely automated, with a minimum of manual work.
Q: Do you plan to develop Desktop app for Atlas?
We don’t plan to implement same functionality with a desktop application. Partially, this functionality is available in photogrammetry processor, which helps stitch maps from images in UgCS Mapper , however, it does not support object detection. If you need object detection, you need to access Atlas cloud installation, or you can get on-premise license and deploy Atlas in your infrastructure.
Q: Does Atlas create the orthomosaics?
Yes, it does.
Q: Can we merge different detectors?
Right now no, but you can have multiple detectors for the same map. You will run them separately, and thus you will generate multiple layers for the same map.