Dear @Frank_van_der_Most and @StephODonnell , thanks for the comments. Regarding the importance of local knowledge in EO data classification, some thoughts follow:

1. Consider two AI applications: large language models (LLM) and object recognition in images.  LLMs such as ChatGPT use words to predict the next word. Since language is its own meta-language, LLMs rely on the fact that our understanding of written text is direct. There are no intermediaries between humans and the printed page. 

2. Object recognition in images (e.g., face recognition) is another kind of AI application where there is an implicit assumption: there are objects (faces, cars, etc) in the image and the role of the algorithm is to distinguish them from the background (considered as unwanted noise).

3. Classification and interpretation of Earth observation data, by contrast, uses a different paradigm. In principle, all of the data is informative. Unlike face recognition, there is no background. Every pixel counts. Pixel values are not words, but  measures of reflections, emissions or echoes of the Earth's surface. 

4. We use words to describe the reality external to us. The variety of nature is such that we have to use simplifications and taxonomies to describe our landscapes. Take the word "forest". As Chazdon et al. question in their 2016 paper, "when is a forest a forest?" The answer is: it depends on who is asking the question. 

5. There have been many attempts to join pixel values with landscape descriptions. e.g, "pixels with NDVI > 0.75 are forests". Do they? What about dry forests that only have high NDVI values in the wet season? So far, all attempts to use direct links between pixel values and landscapes have failed the test of rigour. 

6. Another example is the algorithm used by Global Forest Watch to measure tree cover gain and loss. As explained in the link below, "Not all tree cover is a forest". As GFW acknowledges, their algorithm has problems distinguishing forest from oil-palm plantation and to identify trees in dry forests (see more at  https://research.wri.org/gfr/data-methods?utm_campaign=treecoverloss2021&Limitations#limitations). 

7. Some of you may know the attempt made by FAO to standardize land use and land cover classification using the LCCS ontology. LCCS describes land properties based only on land cover types, disregarding land use. For example, LCCS does not distinguish ‘pasture’ from ‘natural grasslands’; it labels both as herbaceous land cover types. Classification in LCCS has no temporal reference. For a more detailed criticism, see Camara (2020). 

8.  There is no shortage of global land cover and land use maps. While these maps provide a general sense of the global picture, very few (if any) have local significance. As those in the WILDLABS community know, local context matters. My favourite example is the Brazilian Cerrado, an endangered biodiversity hotspot. In the last decades, many areas of natural vegetation in the Cerrado have been converted to pasture for cattle raising. However, global maps inevitably label both pastures and natural Cerrado vegetation as "grasslands". Clearly, such data is hardly usable for supporting studies and public policies in the Cerrado.

9. What is the alternative for mapping areas such as the Cerrado? The only way I see is gathering experts who understand the uniqueness of each ecosystem and try to relate each landscape to signals measured by EO satellites. This is hard and painstaking work, which many iterations. 

10. The recent availability of open big EO data is a blessing and a curse. Using time series, experts can use the temporal evolution of the pixel values to improve the discriminatory power of EO data. Take the distinction between herbaceous pasture and natural Cerrado vegetation. All savannas of the planet (including the Cerrado) have evolved to be resilient to the dry season and to fire. Therefore, while in the wet season it is sometimes difficult to distinguish between herbaceous pasture and natural Cerrado, such distinction increases in the dry season. This is a case of where time series and big data improve the classification results.

11. Big EO data is also a curse, since it requires experts to rethink how to use EO data for land classification. Selecting training samples by looking at a single image is too simplistic when we are classifying time series. Linking the values of a time series to the temporal evolution of the landscapes requires relearning what EO data is. 

Long story short: using Earth observation for conservation studies and public policy making is hard. It requires the combination of big EO data, good algorithms, and lots of expertise to understand the information inherent in the data. A nice challenge to all!

References cited:

Chazdon et al., "When is a forest a forest? Forest concepts and definitions in the era of forest and landscape restoration". Ambio, 45, p 538–550 (2016).

Camara, "On the semantics of big Earth observation data for land classification". Journal of Spatial Information Science, 20 (2020).