A few tips, maybe, from what I have so far realised.

I have had to focus first on validation. A random split can look very strong but fail badly here because the test set is from a different period/distribution. A grouped or spatial/temporal-style validation gave me a much better signal than random CV.

Feature-wise, the biggest value came from robust summaries rather than relying only on raw monthly bands. Raw Sentinel-1/2 monthly values help, but I found temporal aggregates, seasonal min/max/range/std, and water/vegetation indices like NDWI, MNDWI, NDVI-style features more stable. SAR ratios and simple texture/moisture proxies can also help. Though I have been careful with the coordinates or nearest-neighbour style features because they may overfit the public split.

For the 0.5 threshold issue, I have thought in terms of calibration during training rather than post-hoc threshold tuning. Class weighting, balanced objectives, probability calibration, and checking whether the positives count made a big difference for F1 at the fixed cutoff.

A single strong tree model for me did well, but an ensemble of different tree-based models usually gave more stable rankings. The biggest jump for me was better validation first, robust temporal/index features second, calibration/class balance third.

I hope this helps -;)