FaresMallouli/zindi-amini-soil-challenge

13th place (Top 10%) solution for the Zindi Amini Soil Prediction Challenge predicting soil nutrient gaps across Africa using satellite data and ensemble ML models. # Amini Soil Prediction Challenge - 13th Place Solution This repository contains the code for my 13th place finish (Top 10%) in the Amini Soil Prediction Challenge on the Zindi platform. The challenge involved building a model to predict 11 essential soil nutrient gaps for farms across Africa, aiming to empower farmers with data-driven recommendations for a more fertile future. This solution demonstrates a robust machine learning pipeline that effectively handles complex satellite data and successfully overcomes a significant data shift between the training and test sets. ### Top Submissions for Individual Nutrients In addition to the overall 13th place finish, this solution also achieved the best RMSE score for **Iron (Fe)** and **Nitrogen (N)**, as highlighted in the official leaderboard announcement below. ## The Core Challenge: Overcoming Data Shift A key difficulty in this competition was the significant distributional drift between the provided training and test data. As shown below, many of the key predictive features had different statistical properties in the test set compared to the training set. A model trained naively on the training data would not generalize well to the test data, leading to poor performance. *Figure 1: Comparison of feature distributions between the training (blue) and test (red) sets, highlighting a clear data shift.* ## My Winning Strategy My approach was centered on robust feature engineering and a powerful ensemble modeling strategy, with pseudo-labeling as the decisive technique to bridge the data shift. ### Key Methodological Highlights: #### Advanced Feature Engineering: * **Multi-Source Satellite Data:** Integrated and processed extensive Earth Observation data from Landsat 8, Sentinel-1, Sentinel-2, and various MODIS products. * **Temporal Aggregation:** Captured dynamic environmental trends by creating features aggregated over multiple time windows (30, 90, 180, and 365 days). * **Interaction Featu …