Automatic Detection of Ephemeral Gullies Using Orthophotographs: Development and Comparison of Three Pixel-Based Methods in Wallonia (Belgium)

Gully erosion is a significant soil erosion process with potentially severe on-site and off-site impacts. Characterizing and monitoring gullies is essential for accurately predicting their occurrence and effectively preventing their formation and extension. However, creating a gully inventory is both time-consuming and subject to operator bias. Given the low spatial density of gullies in most contexts and their sometimes ephemeral nature, the use of remote sensing appears unavoidable.

In Wallonia (Belgium), gullies mainly occur as ephemeral gullies on agricultural land. Their average width could be as small as 50 cm. High-resolution (25 cm) orthophotographs covering the entire territory are acquired on an annual basis and freely available. Although gullies mostly form on bare or poorly covered soil, depending on the dates of gully formation and image acquisition, gullies may be surrounded by bare or vegetated soil.  

As of today, no methodology has been developed for automatically detecting gullies with such small dimensions at regional scale. This study therefore aims at developing a methodology for automatically detecting ephemeral gullies in Wallonia by remote sensing.

Based on June 2019 orthophotos, 67 gullies across 32 agricultural plots with varying soil cover (including bare soil) were digitized. 16 plots (34 gullies) were used for calibration and 16 plots (33 gullies) for validation. Gully and non-gully pixels were defined inside and immediately outside each gully, respectively, while accounting for delineation uncertainty. An optimal NDVI threshold for each gully was defined as the value maximizing the F1-score of the confusion matrix.

Three pixel-based classification approaches were considered: (1) a fixed NDVI threshold, set to maximize the average F1-score across all gullies of the calibration dataset; (2) a plot-specific NDVI threshold derived from a regression equation linking the median NDVI of the agricultural plots to their optimal NDVI thresholds. This equation was established to optimize the average F1-score (0.82); and (3) a multi-variable Random Forest model.

The use of a fixed NDVI threshold (NDVI = 0.13) achieved an F1-score of 0.77 on the validation dataset. Proper gully detection was achieved only when the median NDVI value of the plot exceeded significantly the threshold value. In contrast, the variable thresholding method allowed to detect gullies even on plots with median NDVI values as low as -0.017 (validation F1-score = 0.82). However, below this value, it failed to detect gullies reliably.

The Random Forest model (3) was trained on the same database. Additional remote sensing as well as topography-related variables were included, such as brightness, the Maximum Difference Index (the maximum difference between the RGB and NIR bands for each pixel), and the distance to the nearest concentrated runoff flow path. Preliminary results are encouraging, and combining this approach with the relative thresholding method could facilitate scaling detection to agricultural plot level and improve post-processing of features with spectral characteristics similar to gullies (e.g., wheel tracks). This approach shows strong potential for large-scale monitoring of gully erosion.