My model reached an F1 score of 0.85 in just 5 epochs, but struggled to improve further beyond that. This rapid early learning suggests that landslide events may have a set of strong, distinguishable features that are easy for the model to pick up. However, the plateau hints at the presence of more subtle or ambiguous cases that are harder to capture. This hypothesis is supported by the optimal threshold I discovered: 0.5. It's unusually high for a problem with significant class imbalance, indicating that the model is relatively confident in its predictions and that true positives are easier to detect than I initially expected.

You can reach 90 on a fold. The larger the model the better

Before switching to the dual-branch EfficientNet architecture, I was using ResNet-18, which gave me my leaderboard score