Land Use/Land Cover Classification Using CatBoost and Multi-Source Remote Sensing Data in the Kurdistan Region of Iraq

• Ragheb Kamal Mohammad1* and Ameer Abdulazeez Abdullah1

  Author

Despite recent advances in remote sensing, classifying land use and land cover (LULC) remains difficult, especially in semi-arid areas, because spectral confusion between classes reduces mapping accuracy. In this paper, we present a multi-stage, multi-source feature-fusion approach for six LULC classes (water, built-up, agriculture, dense forest, open forest, and bareland) using the CatBoost classifier in the intersecting parts of the Amedi, Mergasur, and Akre districts of the Kurdistan Region of Iraq (KRI). Features are ordered into seven different scenarios (from A to G), starting with the visible bands (Scenario A) as baseline and progressing up to adding Sentinel-2 NIR/SWIR bands, land surface temperature (LST), spectral indices (NDVI, NDWI, NDBI, BSI), GLCM textures, Sentinel-1 SAR backscatters, and SRTM topographic predictors (DEM, slope, aspect). We used 10-fold cross-validation repeated 3 times (30 runs per scenario; cross-validation set: 1,200 points), and independently validated the best-performing scenario on 600 held-out test points (100 per class). We compared scenarios using Friedman tests, post-hoc Wilcoxon tests (with Holm correction), and SHAP-based interpretability tests. Overall accuracy (OA) of cross-validation improved from 83.39% (Kappa = 0.801, F1 = 83.26%) in Scenario A to 95.50% (Kappa = 0.946, F1 = 95.48%) in Scenario G, which represents an improvement of 12.11 percentage points (χ²F = 131.11, p < 0.001, n = 30, k = 7). After applying Holm correction, four out of six scenario-pair transitions are statistically significant. The largest improvements came from adding NIR/SWIR bands (+6.14 percentage points) and topographic predictors (+2.31 percentage points). The SHAP and CatBoost feature importance rankings are mostly consistent, ranking slope and elevation as the most important predictors. However, SHAP values reflect a feature's importance in making an individual prediction, not in improving model accuracy. Across 600 independent test points, Scenario G achieved an overall accuracy of 95.33%, a Kappa of 0.944, and a weighted F1-score of 95.35%. These findings demonstrate that integrating multi-source satellite data with an appropriate machine learning method significantly improves LULC classification in semi-arid environments characterised by complex topography.

• pdf