Add context-based post processing for linear features

[rwood-97]

Summary

As per #339, this PR implements a post-processing script so that users can filter out false positives.
This works for linear features or anything where you expect multiple patches to be clustered but solo patches would be false positive.
It also adds a save_predictions() method the classifier to make sure predictions and confidence scores are saved in format expected for post-processing.

Fixes #218
Addresses #339

Checklist before assigning a reviewer (update as needed)

• Allow users to pick lowest conf for which to change label
• Check for edge cases - overlapping patches and non-square patches
• Self-review code
• Ensure submission passes current tests
• Add tests
• Update relevant docs

Reviewer checklist

Please add anything you want reviewers to specifically focus/comment on.

• Everything looks ok?

[codecov-commenter]

Codecov Report

Attention: 11 lines in your changes are missing coverage. Please review.

Comparison is base (b52086e) 59.64% compared to head (f668a73) 60.49%.
Report is 2 commits behind head on main.

❗ Current head f668a73 differs from pull request most recent head 777c857. Consider uploading reports for the commit 777c857 to get more accurate results

Files	Patch %	Lines
mapreader/classify/classifier.py	23.07%	10 Missing ⚠️
mapreader/process/post_process.py	98.50%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #342      +/-   ##
==========================================
+ Coverage   59.64%   60.49%   +0.85%     
==========================================
  Files          35       37       +2     
  Lines        6165     6334     +169     
==========================================
+ Hits         3677     3832     +155     
- Misses       2488     2502      +14     

Flag	Coverage Δ
unittests	60.49% <93.60%> (+0.85%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[rwood-97]

To run, first run inference on some patches and save outputs by calling my_classifier.save_predictions(set_name="infer") with your chosen dataset.

Then:

import pandas as pd

from mapreader.process.post_process import PatchDataFrame

df = pd.read_csv("./predictions_patch_df.csv", index_col=0)

labels_map = {
    0: "no",
    1: "railspace",
    2: "building",
    3: "railspace+building"
}

patches = PatchDataFrame(df, labels_map=labels_map)
patches.get_context(labels=["railspace", "railspace+building",])
patches.update_preds(remap={"railspace": "no", "railspace+building": "building"}, conf=0.8)

This will select all railspace/railspace+building patches, get their context, then update predictions for patches with no surrounding railspace and confidence score of less than 0.8.

Can also set remap to be new labels completely (e.g. {"railspace": "check me", "railspace+building": "check me"}).

[rwood-97]

See here for stats on post-processing https://github.com/Living-with-machines/railspace/issues/14

[edwardchalstrey1]

I guess you could be even more explicit and say: "The current method checks whether any of the 8 surrounding patches have the same label as a given patch (e.g. railspace), and if not, assumes this to be a false positive".

Perhaps could also mention: "Future releases may add functionality to create custom filter rules for your use case"

[edwardchalstrey1]

One comment, but otherwise LGTM