The United Nations (UN) has issued a global appeal to address the challenges associated with land degradation (Goal No. 15), promote sustainable agriculture, and reduce food insecurity (Goal No. 02). The occurrence of gullies significantly contributes to land degradation, which hampers sustainable agriculture, leading to food insecurity. Therefore, it is crucial to understand the spatial distribution of gullies and their related factors in order to achieve the UN's sustainable development goals. Although machine learning (ML) has been successfully applied to gully erosion, there is still considerable untapped potential for further advancement and exploration in this field. Currently, it remains unclear which combination of factors, such as resampling techniques (bootstrapping or k-fold cross-validation), seasons (wet or dry), class numbers (binary or multiclass), and algorithms, yield the highest accuracy, particularly when utilizing low-cost, high-resolution sensors like SPOT-7 and PlanetScope, which have not been extensively researched from the perspective of gully erosion. Additionally, there is a lack of knowledge regarding how varying feature set sizes affect the performance of different algorithms in gully susceptibility modeling. This dissertation addresses these research gaps by employing various ML techniques leveraging low-cost, high-resolution sensors and open-source geospatial data. Multiple accuracy metrics and statistical tests were employed to assess the potential utility of SPOT-7 and PlanetScope images, model gully susceptibility, evaluate ML performance and gain insights into factors influencing accurate gully identification. Results revealed that all SPOT-7 multispectral bands could differentiate gullies, with the near-infrared (NIR) band being the most effective. Moreover, a multiclass approach confirmed models' significance (p < 0.05) in gully identification for each combination of bands and study sites. Regarding PlanetScope imagery, gullies exhibited significantly different (p < 0.001) spectral profiles compared to other land cover classes, both in wet and dry seasons, and NIR was the most influential band. Support Vector Machines (SVM) applied to PlanetScope data demonstrated the best performance with k-fold cross-validation, particularly during the wet season. The classification of gullies was strongly influenced by their characteristics, and the applied algorithms showed the highest efficiency in areas where gullies were deep and exhibited a dendritic pattern. Concerning gully susceptibility modeling, SVM outperformed other algorithms across all overall accuracy metrics, with a relatively short computation time (<1 minute). On the other hand, ensemble-based algorithms, mainly random forest (RF), required a larger set of predictors to achieve maximum accuracy and took several minutes of computation (approximately 15 minutes). Artificial neural networks (ANN) were also sensitive to the number of input features, but unlike RF, their accuracy and computation time consistently decreased with larger feature sets. These techniques can potentially be applied in various regions of the country to support targeted gully rehabilitation and soil management efforts. However, these techniques, particularly satellite image-based gully extraction, are subject to certain limitations and can only be employed in arid or semi-arid regions where gullies are not obscured by dense vegetation, as vegetation obstructs accurate gully extraction. Despite this limitation, this approach offers a solution to the scarcity of laser-scanned data, especially in the study region, by providing improved accuracy and broader application possibilities.