diff --git a/docs/source/examples.md b/docs/source/examples.md index b04b0066d0ec07eac96e8b2ce6304554dc495b81..7a886dd662db4c204bdbf4f944066d4bd4c58a4e 100644 --- a/docs/source/examples.md +++ b/docs/source/examples.md @@ -5,6 +5,8 @@ The goal in this example is to map different forest types using multispectral Sentinel 2 data. +We have a raster and ... points that have been labelled by photo-interpretation. + ```{image} ./_static/examples/original_points.png :alt: Original image with GT points :class: centered-image @@ -12,13 +14,26 @@ The goal in this example is to map different forest types using multispectral Se :align: center ``` -Example table (temp) +Here is a recap of the random forest accuracy given different pre- and post-processing of the encoder's features. + +| Using a DL encoder | Pre-processing | Post-processing | Accuracy | +| :----------- | :----------- | :-------------- |---------: | +| No | No | No | | +| Yes | No | No |0.75 | +| No | 3D PCA | No |**0.93** | +| Yes | 3D PCA | No |0.86 | +| Yes | No | 10D PCA |0.86 | +| Yes | 3D PCA | 10D PCA |0.90 | + -| Left | Centered | Right | -| :----------- | :--------------: | -------------------------: | -| This is left | Text is centered | And this is right-aligned | -| More text | Even more text | And even more to the right | +As we can see, the use of a ViT encoder does not necessarly improve the accuracy of the calssification. +Indeed, the backbone used here was trained on RGB images, working on multispectral data seems to be to far a step here. +Moreover, the different forest types have allready diffrente spectral responses. +By desing, ViT are better than other ML methods to interpret the structure of an image and the relation between patches. +Here, the information being mosty spectral leads to poorer performances after the use of a DL encoder, since the spectral information is drowned between unrelevant structural and tetural information interpreted by the DL backbone. + +Here is the resulting map obtained when infering the RF model without pre-processing: ```{image} ./_static/examples/rf_no_red.png :alt: Random Forest inference without dimension reduction @@ -27,6 +42,8 @@ Example table (temp) :align: center ``` +And with pre-processing: + ```{image} ./_static/examples/rf_red.png :alt: Random Forest inference with dimension reduction :class: centered-image