new chapter on embeddings

Author: gilbertocamara

_quarto.yml

@@ -62,6 +62,9 @@ book:
       chapters: 
         - vec_obia.qmd
         - vec_creating.qmd
+    - part: embeddings.qmd
+      chapters: 
+        - emb_build.qmd
     - part: annex.qmd
       chapters: 
         - annex_export.qmd

emb_build.qmd

@@ -0,0 +1,40 @@
+---
+title: "Object-based time series image analysis"
+format: html
+---
+
+### Configurations to run this chapter{-}
+
+:::{.panel-tabset}
+## R
+```{r}
+#| echo: true
+#| eval: true
+#| output: false
+# load package "tibble"
+library(tibble)
+# load packages "sits" and "sitsdata"
+library(sits)
+library(sitsdata)
+# set tempdir if it does not exist 
+tempdir_r <- "~/sitsbook/tempdir/R/vec_obia"
+dir.create(tempdir_r, showWarnings = FALSE)
+```
+## Python
+```{python}
+#| echo: true
+#| eval: true
+#| output: false
+# load "pysits" library
+from pysits import *
+from pathlib import Path
+# set tempdir if it does not exist 
+tempdir_py = Path.home() / "sitsbook/tempdir/Python/vec_obia"
+tempdir_py.mkdir(parents=True, exist_ok=True)
+```
+:::
+
+
+<a href="https://www.kaggle.com/code/esensing/object-based-image-time-series-classification" target="_blank"><img src="https://kaggle.com/static/images/open-in-kaggle.svg"/></a>
+
+## Introduction

embeddings.qmd

@@ -0,0 +1,16 @@
+---
+title: "Foundational Models and Embeddings"
+format: html
+--- 
+
+## Introduction
+
+The ideas of foundational models and embeddings are among the most recent developments in field of Earth Observation (EO). A Foundational Model (FM) in the context of Earth Observation is a deep learning architecture that is pre-trained on unannotated, multi-modal satellite data. A FM learns the fundamental physical, spatial, and temporal rules of the Earth’s surface by finding patterns in the data. Since does not use labelled samples, this mode of operation is called *self-supervised learning (SSL)*.
+
+The two most common SSL methods used by foundational models are:
+
+* Masked Autoencoders (MAE): The model is fed a satellite image or data cube where a large portion (often up to 75%) of the values has been masked. The model's aims to reconstruct the missing values based on the available values. To do this successfully, the model must learn what features naturally co-occur in the real world. In the case of image patches, the model learns boundaries of objects; when processing  time-series data, the FM learns the temporal signatures of land cover and land use pixels.
+
+* Contrastive Learning: Techniques such as Barlow Twins present the model with two slightly augmented views of the same geographic location (e.g., an image with simulated cloud cover and one without, or time series with different ). The model is forced to mathematically align the representations of these two views while pushing apart the representations of entirely different locations. This teaches the model to ignore atmospheric noise and focus on the underlying biophysical reality.
+
+