1. Introduction

China’s Loess Plateau is the most typical loess geomorphic region in the world and one of the most serious soil erosion regions in China and perhaps the world (Fu et al. , 2011; Zhong et al. , 2022). Gully erosion is an important type of soil erosion on the Loess Plateau, with its distribution area accounting for more than 70% of the ravines and its erosion accounting for 26.6–59.2% of the total slope erosion (Zheng et al. , 2006). Soil erosion control on the Loess Plateau started in 1999 when the Chinese government implemented a project returning farmland to forest (Huang et al. , 2020; Liu et al. , 2019); however, gully erosion still needs attention in some areas. Watershed is the basic unit of hydrological response and an ideal spatial scale for studying soil and water losses. Clarifying the temporal and spatial evolution of gullies at the watershed scale is important for optimizing soil and water conservation measures on the Loess Plateau.
Gully erosion is a linear erosion pattern occurring on steep slope cultivated land, formed by the combined action of runoff erosion and human cultivation, and plays a connecting role in the slope gully erosion system (Liu et al. , 1988; Poesen et al. , 2003; Wang et al. , 2003). At present, studies on gully erosion have focused on critical topographical conditions (Daggupati et al. , 2014; Feng, 2022; Maugnard et al. , 2014; Torri and Poesen, 2014), factors influencing gully formation (Feng, 2022; Gong et al. , 2011; Xu, 2018), mechanical processes and control measures of gully development (Guo, 2019; Xiao, 2017), flow dynamics and sediment transport (Ban et al. , 2020; Kang et al. , 2021a; Xuet al. , 2021), and gully erosion models (Douglas-Mankin et al. , 2020; Guo et al. , 2019; Luquin et al. , 2021; Tekwaet al. , 2021). However, large-area field surveys are needed to accurately grasp the morphological features and spatial distribution characteristics of widely and densely distributed gullies, requiring a large workload and low efficiency. Therefore, some studies have adopted indoor model tests and field slope unit positioning tests to garner relevant gully data (Shen et al. , 2021; Wang et al. , 2020b, 2021), but the limited data significantly impacts the reliability of the research conclusions for practical application. In addition, further verification is needed to determine whether the conclusions obtained under small-scale conditions can be extrapolated to larger scales.
The recent rapid development of remote sensing technology has provided high-resolution remote sensing images and data for gully surveys on a large scale. Some studies have applied remote sensing images for gully recognition (Cao et al. , 2020; Dai et al. , 2020; Liuet al. , 2021; Yu et al. , 2018), determining temporal and spatial variation of gullies (Karydas and Panagos, 2020; King et al. , 2005; Li et al. , 2007; Platoncheva et al. , 2020; Yanet al. , 2005, 2006, 2010; Yermolayev et al. , 2020; Zhonget al. , 2022) and gully erosion sensitivity analyses (Amiriet al. , 2019; Arabameri et al. , 2020; Busch et al. , 2021; Garosi et al. , 2019). On the Loess Plateau, Zhao et al. (2011) analyzed the correlation between land use, slope, and gully distribution using SPOT images, reporting significant differences in gully density among different land use types (grassland > forest land > hilly dry land). Zhang et al. (2017) extracted the lengths of 245 gullies using two QuickBird images, quantified the relationship between gully length and eroded volume, and assessed the erosion rate over six years. Qin (2009) and Qin et al. (2010) analyzed the topographic characteristic parameters and distribution law of gully erosion using a QuickBird image and digital elevation model (DEM), revealing that the morphological features of gullies are determined mainly by slope gradient, slope length, uphill length and confluence area in hilly and gully regions on the Loess Plateau. Wang (2020) found high gully development and large overall erosion potential in the south and north of Dongzhiyuan on the Loess Plateau and low gully development in the central region, where gully erosion is mostly caused by human activities. Liu et al. (2014) and Liu (2012) used 137Cs tracer technology to study the spatial differentiation of soil erosion of typical hilly slopes in hilly and gully areas of the Loess Plateau, revealing differences in hilly slope erosion in different slope directions and an average erosion rates ranked north slope > southwest slope > northeast slope > west slope > northwest slope > south slope > southeast slope > east slope. Jiang et al. (1999) analyzed the distribution law of gullies in the Zhoutungou watershed in the hilly and gully region of the Loess Plateau using aerial images. They found that gullies develop mainly on slope farmland, with gullies accounting for 31.3% of the hillslope area in the watershed. Li (2011) found that the gully density in the loess area of northern Shaanxi, China changes into sunlit slope < semi-sunlit slope < sunless slope < semi-sunless slope under the same slope gradient. Tian et al.(2013) investigated the spatial differentiation characteristics of gullies on the Loess Plateau using 5 m resolution DEM; the results revealed clear spatial differentiation of gully density, decreasing from south to north, with gully density peaking in the Suide-Mizhi area of northern Shaanxi. The above studies were based mainly on visual interpretation and used independent slope elements in the basin as the research object. However, it is not possible to clarify the temporal and spatial distribution and evolution law of gullies at the basin scale. Few studies have identified gullies by combining deep learning and remote sensing images or investigated changes in morphological features before and after returning farmland to forests. Thus, this study used 0.5 m resolution remote sensing images integrated with a deep learning image semantic segmentation model to identify gullies and extract morphological features at the watershed scale, analyze the temporal and spatial distribution of gullies, and clarify the temporal and spatial evolution law of gullies in typical small watersheds in the hilly and gully region of the Loess Plateau to provide a theoretical and scientific basis for watershed-scale gully erosion control and ecological environment protection on the Loess Plateau.