Small fix to doc, replace zeros by Nan before applying offset with S2

Author: loicdtx

docs/deploy.rst

@@ -81,6 +81,7 @@ Here is an example of producing a forest mask using ``xinfereo``:
     )
 
     # 4. Apply standard offset and scaling
+    ds = ds.where(ds != 0)
     ds = (ds.astype(np.float32) - 1000) / 10000
 
     # 5. Initialize the EOInferencer with a pre-trained model
@@ -447,4 +448,4 @@ Considerations
     Filesystem Stability: Depending on the architecture you are deploying on (e.g., HPC, cloud), writing directly to a distributed or network file system can be risky. An interrupted process could leave files in a corrupted or incomplete state. A more stable pattern is to write all outputs to a temporary local disk first (e.g., ``/scratch``). Once all processing is complete, a single, atomic move operation can transfer the final products to the persistent storage.
 
 .. note::
-    Parallel Processing with Dask: For large-scale deployments, the data loading strategy presented above may be inneficient. Dask provides a powerful solution for parallel and out-of-core (lazy) computation. When using ``odc.stac.stac_load``, you can pass a chunks argument (e.g., chunks={'x': 1024, 'y': 1024}). This instructs stac_load to return a Dask array instead of a NumPy array, allowing data to be loaded, processed, and written in parallel blocks.
+    Parallel Processing with Dask: For large-scale deployments, the data loading strategy presented above may be inneficient. Dask provides a powerful solution for parallel and out-of-core (lazy) computation. When using ``odc.stac.stac_load``, you can pass a chunks argument (e.g., chunks={'x': 1024, 'y': 1024}). This instructs stac_load to return a Dask array instead of a NumPy array, allowing data to be loaded, processed, and written in parallel blocks.