Thanks to the organizers. I have endless hours of fun understanding this problem. I am glad to share my approach.

**Detecting Human Settlements**

The competition details can be reviewed here: [INEGI-GCIM Human Settlement Detection Challenge](https://zindi.africa/competitions/inegi-gcim-human-settlement-detection-challenge).

Using the code provided in my GitHub repository (https://github.com/joaquinsalas/zindi), I achieved the following scores:

- **Public leaderboard**: 0.8628

- **Private leaderboard**: 0.8662

These scores compare well with the top of the leaderboard.

---

**Approach**

The solution implements the **Simplified Structured State Space Sequencing (S5)** model, as described by [Smith et al.](https://arxiv.org/abs/2208.04933). The S5 model extends the Structured State Space (S4) ideas of [Gu et al.](https://arxiv.org/abs/2111.00396), which use multi-input, multi-output state space models. S4 itself expands on linear time-invariant dynamical systems (LTI), governed by equations like:

\[

\dot{x}(t) = A x(t) + B u(t), \quad y(t) = C x(t) + D u(t)

\]

Gu et al. condition \(A\) with a low-rank correction, allowing diagonalization, and both papers provide source code.

Additionally, my solution improved based on a suggestion by [Koleshjr](https://zindi.africa/competitions/inegi-gcim-human-settlement-detection-challenge/discussions/23313), who recommended incorporating spectral indices into the model.

---

**Key Insights**

- Use the **S5** model for structured state space learning.

- Balance classes by undersampling the majority class.

- Include spectral indices alongside the image spectral bands for better feature representation.

---

**Instructions**

To replicate my results:

1. Install the S5 model provided by i404788 on GitHub: [S5 PyTorch](https://github.com/i404788/s5-pytorch).

2. The repository includes:

- Code for training the model: `settlements_s5_c17c.py`

- Code for preparing a submission to Zindi: `settlements_s5_c17_zindi.py`

- The best model weights (too large for GitHub; feel free to ask, and I’ll be happy to share).

- The best submission file: `best_s5_model_submission.csv`.

Feel free to ask questions or share feedback. Happy modeling!