Within the realm of Earth Observation, Machine Learning plays a substantial role in facilitating data pinpointing and extraction. The ability to autonomously process and analyze large quantities of data with the help of various techniques within Machine Learning and Computer Vision, has significantly sparked the interest of Earth Observation analysts within the field as of recent years.
Though it may not be readily obvious at first, satellite data imagery processing and analysis brings valuable contribution to a broad range of domains, often beyond Artificial Intelligence. In this regard, numerous use cases may be identified, among which: renewable energy area suitability evaluation, improving disaster response, vegetation and crops monitoring, active conflict areas monitoring and many others.
One might (naively) assume that what the eye can see is all there is to it. In fact, the value of satellite data comes, more often than not, from analyzing parameters outside the visible spectrum of wavelengths. Translating wavelengths that are not visible to the human eye into colors may aid in accurately distinguishing features of interest. For instance, false color infrared greatly emphasizes vegetation in bright red with everything else being colored in darker tones. False color urban band combination may be used to outline urban regions as well as areas with flooding risk. Furthermore, NDWI2 band combination enhances water presence in drought affected areas.
Shifting focus to the technical details, it is commonly acknowledged that “A machine learning model is only as good as the data it is fed” and Earth Observation models are not exempt from this remark. Thus, several procedures are frequently employed to prepare and clean such datasets, including but not limited to: geometric correction, radiometric correction and atmospheric corrections. However, processing data from satellites is no easy feat, considering how large such data tends to be, especially considering spatial and aerial imagery. There are plenty of software tools and open source libraries which are highly specialized in working with large files, among which: BigTiff, Rasterio, Georaster, GDAL and others. Convolutional neural networks are most frequently used with regard to satellite data analysis, though several other architectural types may be employed to aid in analyzing metadata such as recurrent neural networks, self-organizing maps or generative adversarial networks.
To conclude, the demand for real time applications using satellite imagery is substantial and is expected to increase as more industries find advantages in adopting such systems.
Several of the references consulted during research:
- Satellite imagery use cases: https://omdena.com/blog/ai-satellite-imagery/
- Satellite imagery cleanup procedures: https://www.esri.com/arcgis-blog/products/arcgis-pro/imagery/clean-up-your-landsat-imagery-removing-cloud-and-cloud-shadow/
- Image processing techniques: https://www.researchgate.net/publication/315303713_Remote_Sensing_Satellite_Image_Processing_Techniques_for_Image_Classification_A_Comprehensive_Survey
- Handling large files: https://towardsdatascience.com/reading-and-visualizing-geotiff-images-with-python-8dcca7a74510
- Overview of wavelengths and band combinations: https://www.skywatch.com/blog/satellite-image-outputs-and-their-applications
Niciun comentariu:
Trimiteți un comentariu