How to obtain better results when capturing elements on geospatial images?

How to obtain better results when capturing elements on geospatial images?

Leer en Espa?ol

Recently I have received many inquiries and comments about the performance of the GeoAI Plugin.

In summary there is a consensus that the tool is very powerful and promising, however, it is not consistent in its results, i.e. in some cases it will capture automatically and/or assisted (interactive segmentation) approximately 95% of the elements, in other extreme cases, the capture will be less than 10%.

This performance is expected, considering that the SAM (Segment Anything) model was not trained to process geospatial data. In addition, the image format expected by SAM (8-bit RGB images) differs from the format of satellite and drone images, so a transformation is required, which may affect the processing.

In conclusion, the SAM (Segment Anything) model is not yet the definitive and infallible solution for automated extraction on geospatial images, but it can become so. For this we need to understand the conditions required to obtain the best results and if necessary complement it with other algorithms.

In this article I will call geospatial imagery satellite imagery, as well as ortho-imagery generated by drones or airplanes.

What alternatives do we have to obtain better results?

We have a wide variety of options, which can be divided into three main groups, which are briefly described below:

  1. Model adaptation
  2. Adequacy of the image
  3. Complementing or using other techniques

Adequacy of the model

Given the scenario described above, it may seem that the next logical step is to create a new AI (Artificial Intelligence) model with our data; however, this is a complex and demanding task. Before doing so, it is advisable to explore other options.

  1. Exploiting the capabilities of the model The SAM model has multiple configuration options for selection capture and automatic segmentation, not all of which are implemented in GeoAI. It is advisable to fully exhaust these options before proceeding down more complex paths.
  2. Implementing an improved model The release of SAM by META has revolutionized the field of computer vision, especially in the interpretation of medical images. This event has led to the generation of several versions of SAM by organizations and individuals.Testing a free and improved version of SAM can be a more efficient option (fast and with minimal effort), compared to the large investment of training our own model. One of the variants of SAM is the lightweight version recently incorporated in the GeoAI plugin. Likewise, versions of SAM have been generated focused on improving the segmentation of complex elements, with intricate paths, such as HQ SAM. Some of these models have the same architecture as the original SAM, which facilitates their implementation.
  3. Train our own model As mentioned above, these are the most powerful and promising, but demanding alternatives. There are two options for this, with their respective advantages and disadvantages:

  • Train a SAM model from scratch with our data.
  • Adapt the SAM model with our data (fine tuning).

Image adaptation

As I mentioned before, the SAM model can give us excellent results with certain images, an example of this is the following video:

The video shows a very interesting case, since one of the most common failures when segmenting with SAM occurs with elements that extend like crop rows, however, in this image it works perfectly.

It is then worth asking ourselves:

  • How is this image different from the others?
  • Can we adapt the image to obtain better results?

Facing these questions using the scientific method, in an organized and systematic way, will allow us to understand how to achieve better results and in the best case a replicable process flow.

Complement or use other techniques

Deep learning models are not the only option to generate similar processes, there are several segmentation algorithms.

I recently obtained excellent results applying OBIA (Object-Based Image Analysis) with subsequent unsupervised classification on a wine region, while SAM was not useful.

It is not only a matter of implementing other methods, but also of incorporating other relevant data, such as surface models, into the process.

The application of these algorithms does not necessarily require programming or licensing, a wide variety are available in Open Source applications such as SAGA, Grass, Orfeo Tool Box and others.

GeoAI QGIS Plugin Enhancements

The arguments presented in this article allow to identify possible improvements in the capabilities of the GeoAI plugin, soon I will publish an article where I will deepen on this.

Final notes

A significant improvement to the plugin is to incorporate a module with processes that allow to adapt the images to obtain better results in the segmentation, however, first it is necessary to determine what these processes are.

To perform this research I need to have several images, especially high resolution, if you want to collaborate you can send me portions of such images and tell me what you want to extract from them.

Related previous publications

要查看或添加评论,请登录

Luis Eduardo Perez Graterol的更多文章

社区洞察

其他会员也浏览了