The 2280 km long Jinsha River has been blocked at least four times in the past 30 years. A landslide damming hazard chain can endanger communities and infrastructures hundreds of kilometers downstream from the damming site in alpine gorges. Past damming events have resulted in severe consequences, demanding a thorough assessment of damming threats along the entire Jinsha River. This study digitizes the Jinsha River and visualizes its topographic, tectonic, hydrologic, and climate characteristics in detail. A two-stage full-probability method is proposed for assessing the damming threats along this river, making it possible to identify potential damming hotspots and high-priority zones for hazard mitigation. It is found that the upper reach of the Jinsha River poses the greatest damming threat, and the threat level gradually decreases downstream. Approximately 33.4%, 36.7%, 20.5%, and 9.4% of the entire length of the Jinsha River are classified as low, moderate, high, and very high threat levels, respectively. Compared with existing hydropower projects, future projects in the upper reach are more likely to be exposed to landslide damming. We highlight the value of basin-scale spatial threat analysis and envisage that our findings will promote more targeted local-scale risk assessments for potential damming hotspots. These outcomes provide the basis for managing the risks of river damming and hydropower infrastructure along the Jinsha River.
The Jinsha River, the upper reach of the Yangtze River, runs through the eastern margin of the Qinghai-Xizang Plateau. It spans a length of 2280 km, beginning from Yushu City in Qinghai Province and ending in Yibin City, Sichuan Province (Fig. 1). The river covers a drainage area of approximately 480 000 km2 and has a head drop of 3270 m, encompassing rich untapped hydropower resources [1], [2]. With over 100 GW of exploitable water power, the Jinsha River has become home to 22 cascade hydropower projects that have either been constructed or are in the planning phase (Fig. 1(a)). By 2020, the planned total hydropower capacity was 83 GW, making it China’s largest hydropower production base [3], [4], [5]. In addition to dams and hydropower stations, many transportation infrastructures have been built along the river, such as highways (e.g., G214, G215, G317, and G318) and numerous river-crossing bridges (Fig. S1 in Appendix A). The Sichuan-Xizang Railway and Chengdu-Kunming Railway also cross the Jinsha River.
The Indian Plate subducts beneath the Eurasian Plate, causing the Qinghai-Xizang Plateau to rapidly crumple and uplift [6]. Substantial tectonic uplift and fluvial incision have generated high-relief narrow river gorges and steep hillslopes that generate extensive landslides [7], [8], [9], [10]. Large landslides have frequently occurred along the Jinsha River. The 1965 Lannigou landslide (volume > 170 million cubic meters) in Luquan County of Yunnan Province was the worst landslide disaster in the past century; it demolished five villages, caused 444 fatalities, and forced 1004 people to relocate. The formed landslide dam, which was 179 m in height and 2.7 km in length, blocked the Pufu River, a primary tributary of the Jinsha River, for eight months [11], [12]. Influenced by the active Xiaojiang fault, the giant ancient Qiaojia landslide within the Baihetan reservoir occurred with an about 1 billion cubic meters deposition volume [13].
The mainstream Jinsha River has been blocked at least four times in the past 30 years: two successive Baige landslides in 2018, the Suwalong landslide in 1997, and the Huashiban landslide in 1996. Table S1 in Appendix A summarizes the basic information about historical landslide dams along the Jinsha River. The documented damming landslides had volumes of 1.35-536 million cubic meters, dam heights of 35-200 m, and peak dam-breaching discharges of 10 000-82 000 m3∙s−1 (Fig. 2). The two successive Baige landslides in 2018 blocked the Jinsha River twice. The barrier lake submerged Boluo and Jinsha villages upstream, and the peak discharge of the subsequent breaching flood reached 33 900 m3∙s−1, exceeding the 10 000-year flood return period; the flood traveled downstream over 670 km, posing significant threats to communities and cascade hydropower projects downstream [14], [15]. Similar landslide dam hazard chains have been frequently observed in the Jinsha River basin. For instance, the rainfall-induced Tanggudong landslide in 1967 blocked the Yalong River and formed a 170 m high dam with a lake capacity of 680 million cubic meters, generating a peak dam-breaching discharge of 57 000 m3∙s−1 [16], [17]. In 1935, the rainfall-induced Luchedu landslide with a volume of 130 million cubic meters destroyed a village, killed more than 280 people, and blocked the Jinsha River with a 50 m high landslide dam [18]. Approximately (33.5 ± 1.4) thousand years (ka) before present (BP), a moraine landslide was triggered by a strong paleoearthquake and formed a giant landslide dam 3.6 km in length, l.4 km in width, and 105 m in height in the first bend of the Yangtze River [19].
Indeed, landslide damming hazard chains in alpine gorges can inundate upstream areas and flood downstream areas, endangering communities and infrastructure far from the damming site. It is essential to thoroughly assess the potential landslide damming threats and potential elements at risk along the entire Jinsha River.
Efforts have been made to investigate the failure mechanism and dynamic processes associated with individual landslide damming events [19], [20], [21], [22], [23], [24], [25], [26]. The understanding of landslide damming hazards and impacts relies on a handful of case studies [14], [27], [28], [29], but a holistic regional picture of landslide damming threats has yet to be defined. With the boom of cascade hydropower development and dense infrastructure construction along the Jinsha River, a systematic assessment of damming threats along the entire Jinsha River has become essential for risk management and cascading dam safety evaluation [30], [31], [32]. However, a comprehensive understanding of the fundamental characteristics of the Jinsha River remains an open issue, and how to systematically assess damming threats along the entire river has never been determined. Therefore, the objectives of this study are to ① digitize the Jinsha River and quantitatively depict its topographic, tectonic, hydrologic, and climate characteristics; ② propose a two-stage full-probability method to assess the damming threats for the entire river; and ③ identify landslide damming hotspots along the Jinsha River. The findings of this work will enhance the understanding of damming threats on hydropower infrastructures along the Jinsha River and help identify potential damming hotspots for risk mitigation.
2. Jinsha River Basin
The Jinsha River is characterized by high topographic relief, substantial tectonic uplift, strong fluvial incision, active seismic activity, and high climate sensitivity [33]. The river stretches 2280 km, starting from the Batang River confluence (3530 m above sea level (asl)) in Yushu City and ending at the Minjiang River confluence (260 m asl) in Yibin City, with an elevation drop of 3270 m (Fig. 1(a)). The Jinsha River first flows south to Shigu Town along the eastern margin of the Qinghai-Xizang Plateau and then swings northeast to the transition zone of the Yunnan-Guizhou Plateau and the Sichuan Basin. The Jinsha River can be divided into three parts: the upper reach from Yushu City to Shigu Town (1820 m asl) with a length of 974 km, the middle reach from Shigu Town to Panzhihua City (980 m asl) with a length of 558 km, and the lower reach from Panzhihua City to Yibin City with a length of 748 km (Fig. 3(a)). The most significant topographic relief [34] along the Jinsha River is located in the world-famous Tiger Leaping Gorge, which features a valley topographic relief of 3000 m, a length of 16 km, an elevation drop of 200 m, a hillslope gradient of 40°-50°, and a valley floor width of approximately 100 m.
The study area has experienced intense Quaternary activity and frequent earthquakes. Fig. 1(b) shows the lithological map of the study area and its active faults. Siliciclastic sedimentary rocks, mixed sedimentary rocks, and carbonate sedimentary rocks are the most common types of rocks, with area proportions of 30.7%, 30.4%, and 11.4%, respectively. The Archaean to Quaternary strata are exposed, with Triassic, Jurassic, and magmatic rocks being the most widely distributed. Nine active faults intersect the Jinsha River (Fig. 1(b)). Many large historical landslide dams were closely related to these active faults, including the Xuelongnang landslide and Suwalong landslide along the Jinsha River fault zone (F2 in Fig. 1(b)), the Zhaizicun landslide along the Chenghai fault (F8 in Fig. 1(b)), and the Qiaojia landslide along the Xiaojiang fault zone (F9 in Fig. 1(b)) [13], [35], [36].
The key statistics of the Jinsha River include (1110 ± 402) m (i.e., mean ± standard deviation, similarly hereinafter) for valley-to-ridge relief (HV), 27.2° ± 6.4° for the hillslope gradient of the valley flank, (630 ± 662) m for the valley floor width, (779 ± 187) mm for the mean annual precipitation, (0.183 ± 0.042)g for the expected peak ground acceleration (PGA) with a 475-year return period, and (0.027 ± 0.026) km∙km−2 for the fault density. The median valley-to-ridge relief values are approximately 1150 m for the upper reach and middle reach, which are larger than that of the lower reach, with a median elevation difference of 981 m (Fig. 3(b)). The median slope angles of the valley flanks for the upper reach, middle reach, and lower reach are 28.2°, 25.9°, and 27.9°, respectively (Fig. 3(c)). The median valley floor width increases from 282 m (upper reach) to 322 m (middle reach) and further to 582 m (lower reach) (Fig. 3(d)). The mean annual precipitation increases from 495 mm upstream to 1139 mm downstream (Fig. 3(e)). The expected 475-year PGA ranges from 0.1g to 0.3g along the Jinsha River (Fig. 3(f)), reflecting the expected ground motion levels in future earthquakes. The median fault density for the middle reach (0.012 km∙km−2) is much lower than those for the upper reach (0.030 km∙km−2) and lower reach (0.028 km∙km−2) (Fig. 3(g)). The mean annual flow rate monotonically increases from 390 m3∙s−1 (Yushu City) to 1870 m3∙s−1 (Panzhihua City) and further to 4770 m3∙s−1 (Yibin City) (Fig. 3(h)). As a proxy of flow rate, the watershed area monotonically increases from 148 000 to 481 000 m2 (Fig. 3(i)), with a progressive reduction in the channel gradient (Fig. 3(a)).
3. Method and landslide inventory
3.1. Framework for the two-stage full-probability method
The two-stage full-probability method specified in Fig. 4 is adopted to assess the damming threats along the entire Jinsha River. In the first stage, a regional landslide inventory is compiled to develop a landslide damming criterion and establish the joint distribution of landslide characteristics. In the second stage, the Jinsha River is digitized into 1140 2-km-long river stretches characterized by transverse profiles, through which the topographic, tectonic, hydrologic, and climate characteristics along the Jinsha River can be visualized. Then, the landslide damming probability for each river stretch can be computed by considering all possible landslide characteristics. Finally, the river reaches with high damming threats are identified. The proposed method is flexible and can be extended to assess damming threats in other landslide-prone alpine gorges.
3.2. Landslide inventory
This study compiles an inventory of 590 landslides, including 157 damming landslides and 433 non-damming landslides. All the landslides are mapped along large rivers in the eastern margin of the Qinghai-Xizang Plateau, and only those non-damming landslides large enough to potentially block a large river are considered in this study [16], [17], [37]. The spatial distribution of the mapped landslides is shown in Fig. S1 and further summarized in Table S2 in Appendix A. Among the mapped damming landslides, 23 are distributed along the mainstream Jinsha River. The rest of the mapped damming landslides are distributed along other large rivers, such as the Yarlung Zangbo River, Nujiang River, Lancang River, Yalong River, Bailong River Basin, Dadu River, and Minjiang River. A damming landslide is a landslide that completely or partially blocks the river channel and forms a barrier lake [38], [39] (Figs. 5(a)-(f)), while a non-damming landslide does not block the river channel (Figs. 5(g)-(i)). The landslides are mapped through visual interpretation of multitemporal high-resolution satellite images on the Google Earth platform, augmented by relevant literature surveys and field investigations. The landslide mapping method is introduced in Section S1 in Appendix A. Table 1 [40], [41] summarizes the data sources used in this study.
3.2.1. Feature extraction
Fourteen features are considered, including four landslide factors (i.e., landslide area, internal relief, distance to river, and slope gradient), four valley topographic factors (i.e., valley floor width, valley-to-ridge relief, width-depth ratio, and concavity), four triggering factors (annual precipitation, peak ground acceleration, distance to fault, and fault density) (Fig. S2 in Appendix A), one hydrological factor (upstream watershed area), and lithology. Table 2 [41], [42], [43], [44] summarizes the definitions, determination methods, and physical significance of these features.
Fig. 6 presents a schematic diagram illustrating the procedure for extracting the valley topographic factors and landslide factors. A polygon, an initiation point, and a transverse profile are used to depict each landslide (Figs. 6(a) and (b)). The landslide area, including the source and deposition zones, was mapped with a polygon. A point is used to represent the landslide initiation point. In addition, a transverse profile of the landslide is used to extract the valley topographic factors and landslide factors. First, the valley domain is automatically identified with the transverse profile (Fig. 6(c)); the method is introduced in Section S3 in Appendix A). The topography controls the runout path of a landslide. It is assumed that the landslide mass flow direction is primarily along the steepest direction and controlled by gravity, similar to surface water flow. The valley domain within the two mountain ridges adjacent to the valley floor represents the area where a landslide may travel into the river channel. Conversely, a landslide that occurs outside the valley domain cannot travel into the river channel. Within the identified valley domain, four valley topographic factors and four landslide factors are extracted (Figs. 6(d) and (e)). The four triggering factors, the hydrologic factor, and the lithology can be extracted using the methods described in Table 2.
3.2.2. Characteristics of the landslide inventory
All the features can be extracted as described above. Fig. 7 compares the distributions of 13 features for both the damming and non-damming landslides. For the four landslide factors, the landslide area, internal relief, and distance to river for the damming landslides are larger than those for the non-damming landslides (Figs. 7(a)-(c)). In contrast, the slope gradients for the damming landslides are smaller than those for the non-damming landslides (Fig. 7(d)). The patterns for distance to river and slope gradient are less intuitive. The likely reason is that the distance to river is highly correlated with the landslide area and internal relief, and the slope gradient has a high correlation with H/L (the definitions of H and L are shown in Table 2).
For the four valley topographic factors, the valley floor width and valley-to-ridge relief for the damming landslides are smaller than those for the non-damming landslides, while the width-depth ratio and concavity are approximately equal for both the damming and non-damming landslides (Figs. 7(e)-(h)). Among the three triggering factors, the annual precipitation and fault density for the damming landslides are greater than those for the non-damming landslides. The distance to fault for the damming landslides is smaller than that for the non-damming landslides (Figs. 7(i)-(k)), suggesting that the damming landslides may have experienced larger driving forces such as earthquakes, tectonic movements, and precipitation. The watershed area for the damming landslides is smaller than that for the non-damming landslides (Fig. 7(l)), indicating steeper terrains and larger reliefs for the damming landslides. Fig. 7(m) shows the percentages of different rock types among damming and non-damming landslides. Most of the mapped landslides include mixed sedimentary rocks, siliciclastic sedimentary rocks, metamorphic rocks, and carbonate sedimentary rocks. Notably, the percentage of damming landslides of mixed sedimentary rocks is greater than that of non-damming landslides, while the percentage of damming landslides of intermediate volcanic rocks is lower than that of non-damming landslides.
As shown in Fig. 7, the distributions of six features (i.e., landslide area, internal relief, distance to river, valley floor width, and upstream watershed area) exhibit significant skewness. Some variables span multiple orders of magnitude, leading to the model being more sensitive to features with larger value ranges and less sensitive to those with smaller value ranges. By taking the natural logarithm of these six features, their values are brought to a comparable scale, reducing the disparities between different features. This process aims to achieve a more balanced weighting of the features in the model development, enhancing its stability and reliability. Fig. 8 shows the correlation matrix of the 13 features listed in Table 2. The correlations among the three variables (i.e., landslide area, internal relief, and distance to river) are greater than 0.8. The correlation between any two other variables is relatively low, with a correlation coefficient < 0.6. The model selection process will automatically determine the optimal combination of features.
4. Landslide damming criterion
4.1. Model development considering class imbalance
A landslide that moves down the valley flank toward the river channel can either block or not block the river, which is a binary classification problem [45], [46]. Logistic regression is suitable for solving binary problems and has demonstrated exceptional performance in assessing landslides and river-damming [47], [48], [49], [50], [51].
Let P(D = 1|θ, x) denote the damming probability for a landslide, in which x = [x1, x2, …, xn] denotes n selected features, θ = [θ0, θ1, …, θn] is a set of model parameters to be determined, and D is the landslide label (D = 1 denotes damming and D = 0 denotes non-damming). The damming probability can be expressed as:
The number of damming landslides (i.e., 157) is much smaller than the number of non-damming landslides (i.e., 433) in the inventory, which is called class imbalance in the literature [52], [53]. Here, we introduce two weighting factors, wD and wND, for damming and non-damming landslides, respectively. After considering the class imbalance, a weighted likelihood function (L(·)) can be expressed as [54]:
where nD and nND are the total numbers of damming and non-damming landslides, respectively; xi are features of the i-th landslide; and wD = Qp/Qs and wND = (1 - Qp)/(1 - Qs) are weighting factors. Here, Qs is the proportion of damming landslides in the inventory (i.e., Qs = 0.266); Qp is the desired proportion of damming landslides, and Qp = 0.5 is adopted in this study to ensure that damming landslides and non-damming landslides contribute equally during model calibration. The model parameters θ can be calibrated by maximizing Eq. (2).
4.2. Model selection and validation
There are 14 candidate features in the inventory (Table 2). Each possible subset of these features in logistic regression defines a potential model. The inclusion of more variables can increase model flexibility and complexity. The Bayesian information criterion (BIC) is a metric that estimates the tradeoff between the goodness of fit and model complexity [55]. A smaller BIC indicates a better model. We compare all possible combinations of features via the BIC to determine the optimal model, which provides the best compromise between model fit and model complexity [56]. The five models with the lowest BIC values are identified (Table S3 in Appendix A). All these models incorporate lnAL, lnL, lnW, DensF, and lnAW, demonstrating the robustness and reliability of these five features in forecasting landslide damming. The optimal model is selected as the landslide damming criterion:
where [θ0, θ1, θ2, θ3, θ4, θ5] = [16.219, −5.653, 5.863, 1.223, −38.196, 1.185]. Larger values of AL and DensF and smaller values of L, W, and AW promote landslide dam formation.
The fact that HV, R, C, PA, PGA, and lithology are commonly left out by the model selection process (Table S3) indicates that these features are either poor representatives of the underlying physical behavior or are already explained by other features. The role of tectonics in driving landslide damming may be described by fault density, and further inclusion of PGA does not contribute to the increase in predictive performance. In addition, PGA is more likely related to landslide initiation rather than the landslide scale. The lithology is not selected by BIC, probably because the lithological types cannot effectively differentiate between damming and non-damming landslides, making it difficult to determine their influence on damming probability.
To assess the performance of the optimal model, five-fold cross-validation was repeated 200 times to capture fluctuations in the predictive performance. The validation results are presented as distributions instead of a single value. Here, we adopt four metrics to evaluate the prediction performance: ① true positive rate (TPR), which is the proportion of positive data (i.e., damming landslides) that are correctly classified as positive; ② true negative rate (TNR), which is the proportion of negative data (i.e., non-damming landslides) that are correctly classified as negative; ③ accuracy, which is the fraction of data classified correctly; and ④ the area under the receiver operating characteristic curve (AUC). TPR, TNR, and accuracy are computed with a damming probability threshold of 50%. The four metrics are computed and summarized in Fig. 9(a). The medians for TPR and TNR are 0.886, suggesting that the class imbalance problem in the inventory is effectively addressed. The optimal model shows excellent predictive performance, as demonstrated by the median accuracy and AUC values of 0.890 and 0.967, respectively.
The ranking of variable importance in the optimal model can be evaluated by the jackknife test [57]. This test removes one variable at a time and uses the AUC value as an indicator. Since there are five variables in the optimal model, five leave-one-variable-out models can be constructed. The full model (i.e., the optimal model) is used as a benchmark for comparison. After the sequential removal of variables, the variable importance can be ranked by the AUC values. The more the AUC decreases relative to the full model, the greater the importance of the removed variable. A five-fold cross-validation test is conducted and repeated 200 times for each model. The AUC values are then compared in Fig. 9(b). The model with the landslide area removed has the lowest AUC values, indicating that the landslide area is the most important variable in the optimal model, followed by the upstream watershed area, distance to river, fault density, and valley floor width.
5. Damming threats along the Jinsha River
A landslide damming criterion is established based on the inventory in the previous section, which can be used to compute the damming probability given the landslide size and location. However, the exact sizes and locations of future landslides are inherently unpredictable. To this end, the second stage of the two-stage full-probability method shown in Fig. 4 is adopted to assess the damming threats by integrating all possible combinations of landslide sizes and locations. We discretize the Jinsha River into river stretches that are 2 km long, each characterized by a transverse profile. The damming probability of each transverse profile can be computed by considering all the possible combinations of landslide sizes and locations. The damming threats for the entire Jinsha River can then be assessed by aggregating the damming probabilities of all the discretized river reaches.
5.1. Digitization of the Jinsha River
For ease of interpretation and management, the entire Jinsha River is divided into 1140 stretches, each 2 km long. First, for each river stretch, a transverse profile perpendicular to the stream centerline at the middle point of the river stretch is used to depict the valley topography of the corresponding river stretch. Second, the transverse profiles of all the river stretches are digitized, and the valley domains for all the transverse profiles are identified. The valley domain ranges between the two mountain ridges adjacent to the valley floor. This is because the interlaced tributaries outside the valley domain do not contribute to the damming of the mainstream Jinsha River. Finally, the four valley topographic factors within the valley domain are extracted. The four triggering factors, the hydrologic factor, and the lithology are extracted using the midpoints of the stream centrelines.
The values of the valley-to-ridge relief, valley floor width, annual precipitation, fault density, and upstream watershed area along the Jinsha River displayed in Fig. 3 were analyzed in Section 2.
5.2. Damming probability for each river stretch
Eq. (3) can be used to calculate the damming probability for a landslide that has occurred (i.e., the landslide area and initiation point are already known). However, the exact area and initiation point for a future landslide are unknown for a given transverse profile. In this case, the damming probability for a given transverse profile can be calculated using the total probability theorem by integrating Eq. (3) with all possible combinations of landslide area and initiation point (the detailed procedures are introduced in Section S4 in Appendix A). A transverse profile has two sides of valley flanks (i.e., the left and right sides). The landslide damming probability (P(·)) for a transverse profile controlled by the left valley flank and right valley flank can be calculated as
where Dleft = 1 and Dright = 1 indicate that the river is blocked by the landslide initiated from the left valley flank and right valley flank, respectively; P(Dleft = 1|AL, L) and P(Dright = 1|AL, L) represent the damming probabilities controlled by the left and right valley flanks, respectively, conditional upon the landslide area (AL) and distance to river (L). These probabilities can be estimated using Eq. (3); f(AL, L) is the joint probability density function (PDF) of AL and L, which follows a bivariate lognormal distribution using the data in the inventory.
The landslide damming probability of a river stretch blocked by at least one side, PD, can be calculated as follows:
5.3. Assessing damming probability along the Jinsha River
Using Eqs. (3), (4), (5), (6), the damming probability for each river stretch can be computed. Due to the rugged topography characterized by 2 km-spaced transverse profiles, the obtained damming probability can fluctuate somewhat. To pinpoint river reaches with high damming probabilities more precisely, a moving average is adopted to smooth the damming probability. The size of the moving average window can influence the outcome. A larger moving average window will result in a smoother outcome but may lead to loss of detail. After comparing different moving average windows, we find that a 10 km moving average effectively alleviates noise fluctuations in the terrain while preserving sufficient detail (Fig. S3 in Appendix A). The damming probability smoothed by a 10 km moving average along the Jinsha River is shown in Figs. 10(a) and (b).
For ease of interpreting the relative damming threats, all the landslide probability indices are classified into four levels using the Jenks natural break approach, namely, low (PD ≤ 0.09), moderate (0.09 < PD ≤ 0.16), high (0.16 < PD ≤ 0.25), and very high (PD > 0.25). The Jenks natural break approach clusters all the values into four threat levels that minimize the intra-level variance and maximize the inter-level variance. Among the 1140 river stretches, 33.4% are classified as low, 36.7% as moderate, 20.5% as high, and 9.4% as very high. As shown in Fig. 10(a), there is a clear pattern in which river stretches with the same threat level are spatially aggregated.
River reaches with high damming threats are commonly characterized by ① a V-shaped valley with high topographic relief and steep valley flanks, ② a small valley floor width, and ③ a high fault density. For example, the river stretches 116 km upstream of Shigu Town fall into the low threat level, while Longpan-Liangjiaren (the world-famous Tiger Leaping Gorge) belongs to the very high threat level. This is because Tiger Leaping Gorge is characterized by a larger valley-to-ridge relief (1500-2700 m), larger hillslope gradient (40°-50°), smaller valley floor width (100-280 m), and greater fault density (0.057-0.074 km·km−2) than the river stretches 116 km upstream of Shigu Town (Figs. 10(c) and (d)).
Fig. 10(e) shows the PDFs of the damming probabilities for the upper, middle, and lower reaches. The upper reach is associated with the largest damming probability, followed by the middle reach and the lower reach. This pattern could be explained by the fact that the valley floor width and upstream watershed area for the upper reach are the smallest, followed by those for the middle reach and lower reach, while the fault density for the upper reach is the largest, followed by those for the lower reach and middle reach (Fig. 3).
The river reaches with the very high and high threat levels (Fig. 10(a)) include 180-120 km upstream of Gangtuo, 82-32 km upstream of Gangtuo, Boluo-Yebatan, Lawa-Suwalong, 38 km upstream of Xulong-66 km downstream of Xulong, 90-124 km downstream of Xulong, Longpan-Liangjiaren (the world-famous Tiger Leaping Gorge), Ahai-Jin'anqiao, and 0-54 km downstream of Xiluodu. The river reaches with the low threat level (Fig. 10(a)) include 112-96 km upstream of Gangtuo, 116-0 km upstream of Shigu Town, 0-66 km downstream of Longkaikou, 110-0 km upstream of Wudongde, 94 km upstream of Baihetan-Xiluodu, and 112-0 km upstream of Yibin City.
For the hydropower projects along the river, Yebatan, Batang, Xulong, and Jin'anqiao are exposed to very high landslide damming threats. Boluo, Longpan, and Liangjiaren are in river reaches with high damming threats. Among the 11 existing hydropower projects, 1, 0, 8, and 2 correspond to the very high, high, moderate, and low threat levels, respectively, while for the 11 planned (or under construction) hydropower projects, 3, 3, 3, and 2 correspond to the very high, high, moderate, and low threat levels, respectively. Planned hydropower projects are more likely to be exposed to landslide damming than existing hydropower projects. The likely reason for this is that the planned hydropower projects are clustered in the upper reach, while the existing hydropower projects are clustered in the middle or lower reach.
6. Future research directions
Past events have shown that landslide lake outburst floods in alpine gorges can propagate hundreds of kilometers downstream, endangering communities and infrastructure far from the damming site. This raises our concerns about the potential threats posed by the landslide damming hazard chain along the Jinsha River. We highlight the value of basin-scale spatial threat analysis (Fig. 10(a)) and envisage that our findings will promote more targeted local-scale risk assessments for potential damming hotspots. The results obtained in this study can provide practical guidance for managing river damming risks along the Jinsha River.
Future studies should concentrate on targeted local-scale risk assessments and proactive adaptation measures to mitigate the adverse impacts of landslide damming hazards:
(1) Effort is required to quantify the risk of hydropower projects exposed to landslide damming hazard chains. The assessment results presented in this study focus on the spatial locations and do not incorporate a time component, instead providing a long-term average basin-scale threat pattern. A research gap exists between the landslide damming threat and its consequences. Although damming events occur only occasionally, they can cause catastrophic damage. Therefore, it is crucial to accurately characterize the magnitude-frequency curve of landslide damming events over a specified period along the Jinsha River, paving the way for quantitative risk assessment.
(2) It is important to re-evaluate the design floods along the Jinsha River. Currently, the design floods for engineering structures and urban planning purposes are established based on short-term historical meteorological floods, disregarding the extreme floods caused by landslide damming. This oversight leads to an underestimation of flood intensity, exposing infrastructure and communities to unprecedented risks [58]. To address this issue, a systematic investigation is required to evaluate how extreme floods caused by landslide damming amplify design floods [59].
(3) Stress testing can be conducted to evaluate the exposure risk of infrastructure caused by landslide damming hazard chains under various scenarios [60], [61]. The purpose is to determine whether the riverine communities will be flooded, whether the storage capacity of a dam is sufficient to accommodate upstream dam-breaching floods, and whether river-crossing bridges could be destroyed in likely dam-breaching events [62].
(4) Various techniques can be employed for landslide detection and deformation monitoring, including Global Navigation Satellite System (GNSS), optical remote sensing, and interferometric synthetic aperture radar (InSAR) [63], [64], [65], [66]. The geometry of potential landslide dams can be revealed via numerical analysis of potentially unstable hillslopes. The dam-breaching process and downstream flood routing process can be simulated to guide emergency management [14], [67].
(5) Efforts shall be made to develop a basin-scale digital twin, offering forecasting, simulation, and visualization capabilities to inform decision-makers. Additionally, the development of prompt quantitative risk assessment tools can aid in planning effective emergency response strategies [68].
7. Conclusions
The Jinsha River is the most important hydropower clean energy base in China. Systematically assessing landslide damming threats is essential for communities, hydropower development, and infrastructure safety. This study proposes a two-stage full-probability method to assess landslide damming threats along the Jinsha River. The findings are summarized as follows:
(1) A comprehensive inventory, including 157 damming landslides and 433 non-damming landslides, was compiled to develop a regional damming criterion. A model selection process was adopted to make a tradeoff between model complexity and model fit. The model constructed using five features (landslide area, distance to river, valley floor width, fault density, and upstream watershed area) is optimal and achieves exceptional predictive performance. Landslide area is the most important variable in the optimal model, followed by watershed area, distance to river, fault density, and valley floor width.
(2) The entire Jinsha River is digitized and visualized. The key features include (1110 ± 402) m for valley-to-ridge relief, 27.2° ± 6.4° for the hillslope gradient of the valley flank, (630 ± 662) m for the valley floor width, (779 ± 187) mm for the mean annual precipitation, (0.183 ± 0.042)g for the expected peak ground acceleration with a 475-year return period, and (0.027 ± 0.026) km∙km−2 for the fault density. The mean annual precipitation increases from 495 mm upstream to 1139 mm downstream, and the mean annual flow rate monotonically increases from 390 to 4770 m3∙s−1.
(3) The landslide damming probabilities along the Jinsha River are calculated using the proposed two-stage full-probability method. Of the total length of river studied, 9.4%, 20.5%, 36.7%, and 33.4% of the river reaches fall into the very high, high, moderate, and low threat levels, respectively. For the hydropower projects along the river, Yebatan, Batang, Xulong, and Jin'anqiao are exposed to very high damming threats, and Boluo, Longpan, and Liangjiaren are located in river reaches with high damming threats.
(4) Potential landslide damming hotspots are identified and visualized. The upper reach of the Jinsha River faces the greatest threat of river damming, followed by the middle and lower reaches. Newly planned hydropower projects on the upper reach may face greater threats from landslide damming than existing projects. Future efforts are required to conduct quantitative risk assessments on infrastructure exposed to landslide damming hazard chains.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (41941017, U20A20112, and 52025094), the Research Grants Council of the Hong Kong SAR Government (16203720), the NSFC/RGC Joint Research Scheme (N_HKUST620/20 and 42061160480), and the Hetao Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone (HZQB-KCZYB-2020083).
Compliance with ethics guidelines
Shihao Xiao, Limin Zhang, Te Xiao, Ruochen Jiang, Dalei Peng, Wenjun Lu, and Xin He declare that they have no conflict of interest or financial conflicts to disclose.
D.Li, X.X.Lu, X.Yang, L.Chen, L.Lin. Sediment load responses to climate variation and cascade reservoirs in the Yangtze River: a case study of the Jinsha River. Geomorphology, 322 (2018), pp. 41-52
[2]
J.He, L.Zhang, T.Xiao, C.Chen. Emergency risk management for landslide dam breaks in 2018 on the Yangtze River China. Resilient Cities Struct, 1 (3) (2022), pp. 1-11
[3]
National Energy Administration. The 13th Five-Year Plan for hydropower development (2016-2022). National Energy Administration, Beijing (2016) Chinese
[4]
F.Jin, D.Huang, M.Lino, H.Zhou. A brief review of rock-filled concrete dams and prospects for next-generation concrete dam construction technology. Engineering, 32 (2023), pp. 99-105
[5]
Z.Bao, J.Zhang, Y.Lian, G.Wang, J.Jin, Z.Ning, et al. Changes in headwater streamflow from impacts of climate change in the Qinghai-Xizang Plateau. Engineering, 34 (2023), pp. 133-142
[6]
M.Hao, J.T.Freymueller, Q.Wang, D.Cui, S.Qin. Vertical crustal movement around the southeastern Qinghai-Xizang Plateau constrained by GPS and GRACE data. Earth Planet Sci Lett, 437 (2016), pp. 1-8
[7]
R.Jiang, L.Zhang, D.Peng, X.He, J.He. The landslide hazard chain in the tapovan of the Himalayas on 7 February 2021. Geophys Res Lett, 48 (17) (2021), Article e2021GL093723
[8]
D.Peng, L.Zhang, R.Jiang, S.Zhang, P.Shen, W.Lu, et al. Initiation mechanisms and dynamics of a debris flow originated from debris-ice mixture slope failure in southeast Xizang China. Eng Geol, 307 (2022), Article 106783
[9]
O.Marc, N.Hovius, P.Meunier, T.Uchida, S.Hayashi. Transient changes of landslide rates after earthquakes. Geology, 43 (10) (2015), pp. 883-886
[10]
J.N.Jones, S.J.Boulton, M.Stokes, G.L.Bennett, M.R.Whitworth. 30-year record of Himalaya mass-wasting reveals landscape perturbations by extreme events. Nat Commun, 12 (1) (2021), p. 6701
[11]
X.Cheng, C.Zhu, W.Qi, J.Xu,J. Yuan. Formation conditions, development tendency, and preventive measures of Pufu landslide in Luquan of Yunnan. Miner Resour Geol, 29 (2015), pp. 395-401Chinese
[12]
B.Wen, F.Wang. Precursors and motion characteristics of the 1965 Lannigou rockslides and the subsequent evolution. Hydrogeol Eng Geol, 48 (6) (2021), pp. 72-80Chinese
[13]
Z.Feng, Z.H.Wu, J.W.Cao, M.M.Hu, G.H.Ha, H.X.Kang. Engineering geological characteristics of gigantic pre-historic landslide along Qiaojia section of the Xiaojiang Fault. Acta Geosci Sin, 4 (2019), pp. 629-636Chinese
[14]
L.M.Zhang, T.Xiao, J.He, C.Chen. Erosion-based analysis of breaching of Baige landslide dams on the Jinsha River, China, in 2018. Landslides, 16 (10) (2019), pp. 1965-1979
[15]
ZhangLM, HeJ, XiaoT. Engineering risk mitigation for landslide hazard chains:the Baige landslides on the Jinsha River in 2018. In: Proceedingsof the 5th World Landslide Forum; Nov2-62021; Kyoto, Japan. Berlin:Springer; 2021.
[16]
H.Chai, H.Liu, Z.Zhang. The catalog of Chinese landslide dam events. J Geol Hazards Environ Preserv, 6 (4) (1995), pp. 1-9Chinese
[17]
Y.Cai, Y.Luan, W.Peng, J.Li, Y.Xu. Analysis on typical cases of barrier lake:causes, disaster and emergency response. Changjiang Press, Wuhan (2021)Chinese
[18]
X.Liu, J.Qiao, N.Li. Landfall-landslide disaster in Yunnan Province and the preventive measures against them. Bull Soil Water Conserv, 7 (6) (1987), pp. 38-44Chinese
[19]
Y.Li, J.Chen, J.Yan, F.Zhou, Q.Wang, Z.Li, et al. Formation and evolution of a giant old deposit in the First Bend of the Yangtze River on the southeastern margin of the Qinghai-Xizang Plateau. Catena, 213 (2022), Article 106138
[20]
Y.Li, J.Chen, F.Zhou, S.Song, Y.Zhang, F.Gu, et al. Identification of ancient river-blocking events and analysis of the mechanisms for the formation of landslide dams in the Suwalong section of the upper Jinsha River, SE Qinghai-Xizang Plateau. Geomorphology, 368 (2020), Article 107351
[21]
Y.Zhang, R.Wu, C.Guo, L.Wang, X.Yao, Z.Yang. Research progress and prospect on reactivation of ancient landslides. Adv Earth Sci, 33 (7) (2018), pp. 728-740Chinese
[22]
C.Guo, Y.Zhang, X.Li, S.Ren, Z.Yang, R.Wu, et al. Reactivation of giant Jiangdingya ancient landslide in Zhouqu County, Gansu Province China. Landslides, 17 (1) (2020), pp. 179-190
[23]
L.Wu, H.Deng, R.Huang, L.M.Zhang, X.Guo, Y.Zhou. Evolution of lakes created by landslide dams and the role of dam erosion: a case study of the Jiajun landslide on the Dadu River China. Quat Int, 503 (2019), pp. 41-50
[24]
P.Cui, Y.Zhu, Y.Han, X.Chen, J.Zhuang. The 12 May Wenchuan earthquake-induced landslide lakes: distribution and preliminary risk evaluation. Landslides, 6 (3) (2009), pp. 209-223
[25]
M.Peng, L.Zhang. Dynamic decision making for dam-break emergency management—part 2: application to Tangjiashan landslide dam failure. Nat Hazards Earth Syst Sci, 13 (2) (2013), pp. 439-454
[26]
N.Liu, Z.Chen, J.Zhang, W.Lin, W.Chen, W.Xu. Draining the Tangjiashan barrier lake. J Hydraul Eng, 136 (11) (2010), pp. 914-923
[27]
P.Cui, G.G.Zhou, X.Zhu, J.Zhang. Scale amplification of natural debris flows caused by cascading landslide dam failures. Geomorphology, 182 (2013), pp. 173-189
[28]
X.Fan, A.Dufresne, S.S.Subramanian, A.Strom, R.Hermanns, C.T.Stefanelli, et al. The formation and impact of landslide dams—state of the art. Earth Sci Rev, 203 (2020), Article 103116
[29]
Q.Zhong, L.Wang, S.Chen, Z.Chen, Y.Shan, Q.Zhang, et al. Breaches of embankment and landslide dams-state of the art review. Earth Sci Rev, 216 (2021), Article 103597
[30]
W.Schwanghart, M.Ryan, O.Korup. Topographic and seismic constraints on the vulnerability of Himalayan hydropower. Geophys Res Lett, 45 (17) (2018), pp. 8985-8992
[31]
L.Berga. The role of hydropower in climate change mitigation and adaptation: a review. Engineering, 2 (3) (2016), pp. 313-318
[32]
F.Lempérière. Dams and floods. Engineering, 3 (1) (2017), pp. 144-149
[33]
P.Cui, Y.Ge, S.Li, Z.Li, X.Xu, G.G.Zhou, et al. Scientific challenges in disaster risk reduction for the Sichuan-Xizang Railway. Eng Geol, 309 (2022), Article 106837
[34]
W.Schwanghart, D.Scherler. TopoToolbox 2-MATLAB-based software for topographic analysis and modeling in Earth surface sciences. Earth Surf Dynam, 2 (1) (2014), pp. 1-7
[35]
J.Chen, F.Dai, T.Lv, Z.Cui. Holocene landslide-dammed lake deposits in the Upper Jinsha River, SE Qinghai-Xizang Plateau and their ages. Quat Int, 298 (2013), pp. 107-113
[36]
Z.Xu, W.Liu, R.Huang. Engineering geological characteristics of Zhaizicun giant ancient landslide along Jinsha River and its occurrence mechanisms. Chin J Rock Mech Eng, 30 (S2) (2011), pp. 3539-3550Chinese
[37]
W.Liu, P.A.Carling, K.Hu, H.Wang, Z.Zhou, L.Zhou, et al. Outburst floods in China: a review. Earth Sci Rev, 197 (2019), Article 102895
[38]
J.E.Costa, R.L.Schuster. The formation and failure of natural dams. Geol Soc Am Bull, 100 (7) (1988), pp. 1054-1068
[39]
O.Korup. Recent research on landslide dams—a literature review with special attention to New Zealand. Prog Phys Geogr Earth Environ, 26 (2) (2002), pp. 206-235
[40]
J.Hartmann, N.Moosdorf.The new global lithological map database GLiM: a representation of rock properties at the Earth surface. Geochem Geophys Geosyst, 13 (12) (2012), pp. 1-37
[41]
X.Xu, Z.Han, X.Yang, S.Zhang, G.Yu, B.Zhou, et al. Seismotectonic map in China and its adjacent regions. Seismological Press, Beijing (2016)Chinese
[42]
K.A.Fryirs, J.M.Wheaton, G.J.Brierley. An approach for measuring confinement and assessing the influence of valley setting on river forms and processes. Earth Surf Processes Landforms, 41 (5) (2016), pp. 701-710
[43]
C.Xu, X.Xu, G.Yu.Landslides triggered by slipping-fault-generated earthquake on a plateau: an example of the 14 April 2010, Ms 7.1, Yushu, China earthquake. Landslides, 10 (4) (2013), pp. 421-431
[44]
S.Qi, Q.Xu, H.Lan, B.Zhang, J.Liu.Spatial distribution analysis of landslides triggered by 2008.5.12 Wenchuan Earthquake, China. Eng Geol, 116 (1,2) ( 2010), pp. 95-108
[45]
X.Fan, D.G.Rossiter, C.J. vanWesten, Q.Xu, T.Görüm. Empirical prediction of coseismic landslide dam formation. Earth Surf Processes Landforms, 39 (14) (2014), pp. 1913-1926
[46]
C. TacconiStefanelli, N.Casagli, F.Catani. Landslide damming hazard susceptibility maps: a new GIS-based procedure for risk management. Landslides, 17 (7) (2020), pp. 1635-1648
[47]
C.Y.Chen, J.M.Chang. Landslide dam formation susceptibility analysis based on geomorphic features. Landslides, 13 (5) (2016), pp. 1019-1033
[48]
H.Wang, L.Zhang, H.Luo, J.He, R.W.M.Cheung. AI-powered landslide susceptibility assessment in Hong Kong. Eng Geol, 288 (2021), Article 106103
[49]
T.Xiao, L.M.Zhang, R.W.M.Cheung, S.Lacasse. Predicting spatio-temporal man-made slope failures induced by rainfall in Hong Kong using machine learning techniques. Geotechnique, 73 (9) (2022), pp. 749-765
[50]
J.Jin, G.Chen, X.Meng, Y.Zhang, W.Shi, Y.Li, et al. Prediction of river damming susceptibility by landslides based on a logistic regression model and InSAR techniques: a case study of the Bailong River Basin China. Eng Geol, 299 (2022), Article 106562
[51]
S.H.Xiao, L.M.Zhang, T.Xiao, R.C.Jiang. Prediction of landslide dam formation using machine learning techniques. Geo-Risk 2023, American Society of Civil Engineers, Reston (2023), pp. 41-48
[52]
J.Zhang, T.P.Wang, S.H.Xiao, L.Gao. Chinese code methods for liquefaction potential assessment based on standard penetration test: an extension. Soil Dyn Earthquake Eng, 144 (2021), Article 106697
[53]
J.Zhang, S.H.Xiao, H.W.Huang, J.K.Zhou. Calibrating a standard penetration test based method for region-specific liquefaction potential assessment. Bull Eng Geol Environ, 79 (10) (2020), pp. 5185-5204
[54]
C.H.Juang, S.Khoshnevisan, J.Zhang. Maximum likelihood principle and its application in soil liquefaction assessment. K.K.Phoon, J.Y.Ching (Eds.), Risk and reliability in geotechnical engineering, CRC Press, New York City (2015), pp. 181-219
[55]
J.Zhang, H.Huang, C.Juang, W.Su. Geotechnical reliability analysis with limited data: consideration of model selection uncertainty. Eng Geol, 181 (2014), pp. 27-37
[56]
A.E.Raftery. Bayesian model selection in social research. Sociol Methodol, 25 (1995), pp. 111-163
[57]
A.R.Swan, M.Sandilands. Introduction to geological data analysis. Wiley-Blackwell, Boston (1995)
[58]
W.Schwanghart, R.Worni, C.Huggel, M.Stoffel, O.Korup. Uncertainty in the Himalayan energy-water nexus: estimating regional exposure to glacial lake outburst floods. Environ Res Lett, 11 (7) (2016), Article 074005
[59]
H.Apel, B.Merz, A.Thieken. Influence of dike breaches on flood frequency estimation. Comput Geosci, 35 (5) (2009), pp. 907-923
[60]
L.Devitt, J.Neal, G.Coxon, J.Savage, T. Wagener. Flood hazard potential reveals global floodplain settlement patterns. Nat Commun, 14 (1) (2023), p. 2801
[61]
J.He, Y.Qiang, H.Luo, S.Zhou, L.Zhang. A stress test of urban system flooding upon extreme rainstorms in Hong Kong. J Hydrol, 597 (2021), Article 125713
[62]
X.Zhou, Z.Chen, J.Zhou, X.Guo, X.Du, Q.Zhang. A quantitative risk analysis model for cascade reservoirs overtopping: principle and application. Nat Hazards, 104 (1) (2020), pp. 249-277
[63]
J.Yao, H.Lan, L.Li, Y.Cao, Y.Wu, Y.Zhang, et al. Characteristics of a rapid landsliding area along Jinsha River revealed by multi-temporal remote sensing and its risks to Sichuan-Xizang Railway. Landslides, 19 (3) (2022), pp. 703-718
[64]
S.Zhou, B.Chen, H.Lu, Y.Shan, Z.Li, P.Li, et al. Analysis of the spatial distribution and deformation types of active landslides in the upper Jinsha River, China, using integrated remote sensing technologies. Remote Sens, 16 (1) (2023), p. 100
[65]
W.Yang, L.Liu, P.Shi. Detecting precursors of an imminent landslide along the Jinsha River. Nat Hazards Earth Syst Sci, 20 (11) (2020), pp. 3215-3224
[66]
C.Zhao, Y.Kang, Q.Zhang, Z.Lu, B.Li. Landslide identification and monitoring along the Jinsha River catchment (Wudongde reservoir area), China, using the InSAR method. Remote Sens, 10 (7) (2018), p. 993
[67]
C.Chen, L.Zhang, T.Xiao, J.He. Barrier lake bursting and flood routing in the Yarlung Tsangpo Grand Canyon in October 2018. J Hydrol, 583 (2020), p. 12460
PDF
(8293KB)
