From a3d09af1c7964e1bc44804840b6b02dfb2d23564 Mon Sep 17 00:00:00 2001 From: Daniel Wiesmann Date: Wed, 1 Mar 2023 10:26:42 +0000 Subject: [PATCH] intermediate --- _posts/2023-02-28-our-models.md | 36 ++++++++++++++------------------- 1 file changed, 15 insertions(+), 21 deletions(-) diff --git a/_posts/2023-02-28-our-models.md b/_posts/2023-02-28-our-models.md index 1dc23c9..471cff5 100644 --- a/_posts/2023-02-28-our-models.md +++ b/_posts/2023-02-28-our-models.md @@ -8,16 +8,21 @@ description: "An introduction to Tesselo's AI modeling, explaining the model typ featured: false hidden: false --- -Tesselo's deep learning models are presented in this post. We have used them to -do large scale land cover modeling across the world. +Tesselo's most successful deep learning models are presented in this post. We have +used them to do large scale land cover modeling across the world. -We have packaged our models into a repository that makes it easy -to use Tesselo's most common models. You can find the model references -in our [Alquimodelia](https://github.com/tesselo/alquimodelia) repository. +We have packaged our most common models into a repository that makes it easy +to use them. You can find the model references in our +[Alquimodelia](https://github.com/tesselo/alquimodelia) repository. It contains +the detailed model definitions for our most successful models. We used Keras with a +Tensorflow backend for our modeling, so the definitions are written in that famework. -The aim of Alquimodelia was to create an user friendly and easy way to use and -change parameters on the common model architectures used in Tesselo. No need of any -knowledge in keras or tensorflow, just some parameters and you had your model ready to use. +The aim of Alquimodelia is to provide a user friendly way to use and change parameters +on the common model architectures used in Tesselo. The model classes can be created without +deep knowledge of keras or tensorflow. The main required parameters are the input and ouput +shape that the models will work with. Then, Arquimodelia will construct the models accordingly. + +## Model types Depending on the context and the goal of the modeling, we have used a series of different models. They range from pixel based classifiers to time-series based @@ -25,18 +30,6 @@ U-Net type architectures. ## Use all bands -For our modeling, we moslty used all available bands of the multispectral satellite -images. For Sentinel-2 we used the 10 bands that have 10m or 20m resolution. Similarly, -for Landsat we used the available bands. - -In our pre-processing pipeline we simply resampled all bands into the target resolution. -Usually this meant to upsample the lower resolution bands to the resolution of the -band with the highest resolution. That is 10m for Sentinel-2 images for instance. - -Or using the same approuch we would create super-resolution, by upsampling our imagery data -to the resolution of the target data. We had successful models that would build 1m resolution -images out of 10m resolution data. - ## Classifiers Here we are giving a quick overview of the different model types and their use cases. @@ -68,6 +61,7 @@ Used in image classification and segmentation. #### 3D U-Net -The 3D U-Net architeture follows the same patterns and the 2D, but instead of two-dimensional convolutions it uses three-dimensional convolutions, multiple images across time. +The 3D U-Net architeture follows the same patterns and the 2D, but instead of two-dimensional +convolutions it uses three-dimensional convolutions, multiple images across time. The answer would still be a single image, but produced with time context. Great to surpass problems like clouds and other imagery artifacts.