Mapping landslides using NDVI variation and slope in Google Earth Engine: A case of study of Santa Tereza municipality due to the most extensive disaster in Brazil (2024)

Brazil has experienced an increase in landslides occurrences associated with extreme rainfall events. In the Taquari-Antas Basin, southern Brazil, the valley-shaped relief favor the development of different types of landslides. During the 2024 extreme rainfall-induced event, more than 16,000 landslide scars were mapped across the state, including 281 in the municipality of Santa Tereza, which presents the highest landslide scar density per area. However, landslides inventories are still largely based on visual interpretation of satellite imagery, manual delimitation, and, when feasible, field validation.

Satellite imagery plays a fundamental role in landslide mapping, particularly in hard-to-reach areas and during disaster events. To improve landslide detection, this study proposes a simple approach combining the variation of Normalized Difference Vegetation Index (dNDVI) and terrain slope, using Sentinel-2 imagery with 10 m spatial resolution. Data processing was performed using Google Earth Engine (GEE).

The dNDVI, calculated from pre- and post-event images, enables the identification of vegetation loss, which is particularly effective in Santa Tereza, where more than 70% of the municipality is forest-covered. As landslides predominantly occur on steep hillslopes in Brazil, slope information was incorporated to refine the detection. The combined analysis of dNDVI and slope resulted in an initial landslide detection map.

NDVI values range from 0 to 1, with higher values indicating denser vegetation. In southern Brazil, low dNDVI thresholds (e.g., 0.10) may misclassify cloud shadows or crop harvesting as landslides, whereas high thresholds (e.g., 0.40) may capture only the core of the scar. A sensitivity analysis was conducted by testing three dNDVI thresholds (0.25, 0.20, and 0.15) combined with three slope thresholds (15°, 10°, and 8°).

Validation was performed by comparing the detection results with a detailed landslide inventory produced by the Latitude/UFRGS research group, classifying the outcomes as true positives, false negatives, and false positives.

The results show true positive rates ranging from 59% to 83%. The best overall performances were the combinations dNDVI ≥ 0.15 with slope ≥ 15°, dNDVI ≥ 0.20 with slope ≥ 15°, and dNDVI ≥ 0.25 with slope ≥ 15°. False negative rates were lowest for dNDVI ≥ 0.15 with slope ≥ 15° combination. False positive rates ranged from 72% to 87%, with lower values observed for combinations of higher dNDVI and slope thresholds. The proposed approach provides a practical and rapid technique to support landslides mapping and post-disaster monitoring.

Acknowledgements: This study was supported by FAPERGS under Grant Agreement No. 24/2551-0002124-8 (Call FAPERGS 06/2024) and No. 25/2551-0002522-2 (Call FAPERGS 05/2025).