Thales Sehn Körting

Deep Learning articles use benchmarks to measure the quality of the results. However, several benchmarks do not have the copyright of all data used. So, how to believe that every paper uses the same benchmark?

Several authors rely on transfer learning from pretrained models, arguing that using well-known datasets, which are available on the internet (e.g. ImageNet) their model will be able to handle a specific problem with a reduced training step.

In Remote Sensing this perspective is also becoming a trend when using Deep Learning techniques to classify Remote Sensing datasets.

In my opinion, the datasets used for pretrain are very different from Remote Sensing targets, mainly in two aspects:

• spatial resolution: a sensor can be ultra high spatial resolution (50cm for example) or very low resolution (2km for a single pixel), and the edges in all these images are different
• spectral resolution: the datasets found on the internet are composed by color pictures, obtained mainly by phone cameras, which are composed by 3 channels (red, green and blue). In Remote Sensing we can have several spectral channels, such as yellow or red-edge bands (available in WorldView-2), or infra-red channels, available in most of the satellites. How to train a model using 3 bands, when in reality you can have at least 5 bands with so different information?

If you agree, or if you do not agree, please give some feedback and let's learn together.

In this podcast I discuss the (sometimes) wrong use of the term Data Mining, with in accord to the paper

From Data Mining to Knowledge Discovery in Databases, written in 1996 by Usama Fayyad, Gregory Shapiro, and Padhraic Smyth, 

is defined as: Data mining is a step in the KDD process that consists of applying data analysis and discovery algorithms that produce a particular enumeration of patterns (or models) over the data.

KDD means Knowledge Discovery in Databases, and is composed by the following steps:

Data -> (selection) -> Target Data -> (preprocessing) -> Preprocessed Data -> (transformation) -> Transformed Data -> (data mining) -> Patterns -> (interpretation/evaluation) -> Knowledge

Several authors call Data Mining when they are performing the entire cycle (from Data to Knowledge) and not only the data mining step, which can be represented also by the use of classification/clustering algorithms.

In this podcast I discuss the wrong use of the term Resolution in scientific articles or in the general media. Resolution in Remote Sensing can be used to describe several aspects of images, such as:

• temporal resolution: the time difference between two images of the same place
• spectral resolution: related to the number of bands and wavelengths, such as in Panchromatic, Multispectral, Hyperspectral, or Ultraspectral
• radiometric resolution: the number of bits needed to store a pixel value (e.g. 8 bits in Landsat 7 or 11 bits in WorldView-2)
• spatial resolution: the focus of this podcast, relating the area represented by a single pixel in an image

I provide an interesting reference with an easy to use table, to understand what can be considered High Spatial Resolution, or Low Spatial Resolution:

Taxonomy of Remote Sensing Systems - Spatial Ground Resolution

• Ultra High: < 1m
• Very High: [1m, 4m]
• High: [4m, 10m]
• Medium: [10m, 50m]
• Low: [50m, 250m]
• Very Low: > 250m

The reference is:

Ehlers, M., Janowsky, R., Gähler, M., 2001. New remote sensing concepts for environmental monitoring. Proceedings of SPIE - The International Society for Optical Engineering.

This is a first message to check if someone will find my podcast and will have interest on it.  Waiting for feedback on remote sensing, image processing, data mining, deep learning, data augmentation, sample selection, articles, papers, etc.