南水北调与水利科技（中英文）2026，Vol.24Issue(4)：1032-1046,15.DOI:10.13476/j.cnki.nsbdqk.2026.0095
基于LSTM地形迁移学习的湖泊水储量演化异质性
Heterogeneity in water storage evolution of lakes based on LSTM bathymetry transfer learning
摘要
Abstract
The Inner Mongolia Plateau's lakes have significantly shrunk due to human activity and climate change,endangering northern China's ecological security.Most long-term assessments used lake area as a proxy for water storage because it was easily obtained through remote sensing.However,basin morphology governed the conversion of water level,surface area,and stored volume,allowing area and water storage to become decoupled,and relying solely on area misrepresented the evolution of water resources.Furthermore,lakes with different morphologies showed noticeably different shrinkage behavior under generally similar anthropogenic and climatic conditions.High-precision underwater topography modeling,however,was still mostly limited to the scale of individual lakes.As a result,the mechanism by which lake evolution responded to basin topography remained poorly constrained,and a regional,morphology-resolved understanding of water-storage evolution was still lacking. A previously published model for predicting underwater topography based on a long short-term memory network(LSTM)was expanded from a single lake to twenty-seven medium-to-large lakes across the plateau.The large differences in lake size and the uneven density of field measurements were reconciled using a two-stage approach that combined global pre-training with per-lake ensemble residual learning.Multi-source data were integrated,including long-term satellite imagery for the extraction of annual water-area series from 1990 to 2024,a global elevation model corrected with satellite laser altimetry,and field bathymetric soundings together with surveyed shoreline elevations.The three-dimensional underwater topography of each lake was reconstructed,and annual water-storage series were calculated by combining the water-surface boundary with the predicted basin floor.A dual-threshold classification framework based on mean water depth and the ratio of the semi-major axis to mean depth was built after morphological parameters were extracted within a uniformly defined maximum stable boundary.Evolution rates were characterized by ordinary least-squares slopes,breakpoints were identified and verified by trend and structural-change tests,and differences among groups were examined with non-parametric tests. Leave-one-out cross-validation produced a mean coefficient of determination of 0.758,and an independent hold-out validation for Daihai yielded 0.761,confirming that the method could be applied to lakes with sparse bathymetric data.The twenty-seven lakes were divided into wide-shallow-open,transitional-mixed,and deep-bowl-compact types;among them,twelve were wide-shallow-open,seven were transitional-mixed,and eight were deep-bowl-compact.The area-to-water-storage rate ratio(V/A)showed an overall difference among the three types that was statistically significant(P=0.001 6),and mean water depth was significantly and positively correlated with V/A(rs=0.78,P<0.001).These findings indicated that vertical basin depth was the most important morphological factor influencing the relationship between area and water storage.For wide-shallow-open lakes,substituting the rate of area change for the rate of water-storage change overestimated the actual change in water volume by more than sixfold.Basin morphology also modulated the consistency of shrinkage,the timing of breakpoints,and the magnitude of interannual fluctuation.Specifically,wide-shallow-open lakes responded earliest to climatic signals and exhibited the strongest amplification of area fluctuation,whereas deep-bowl-compact lakes shrank most consistently and responded most stably. Basin morphological structure was identified as the key internal factor that controlled the heterogeneity of lake evolution on the Inner Mongolia Plateau.Water storage,rather than area,was shown to be a more reliable and more universal indicator of lake evolution,particularly for wide-shallow-open lakes,for which area-based assessment could substantially overstate the magnitude of change.Basin morphological type should be incorporated as an essential reference dimension in the analysis of lake evolution and in the classified management of water resources.A monitoring and early-warning scheme based on water storage should be given priority in wide,shallow,open lakes so that changes in surface area are not mistaken for the true variation of water resources.关键词
内蒙古高原/湖盆形态/双阈值分类/水储量/面积演化差异Key words
Inner Mongolia Plateau/lake basin morphology/dual-threshold classification/water storage/difference in area evolution分类
建筑与水利引用本文复制引用
蒋瑞瑞,马荣,东王刚,焦梦青,赵乐凡..基于LSTM地形迁移学习的湖泊水储量演化异质性[J].南水北调与水利科技（中英文）,2026,24(4):1032-1046,15.基金项目
国家重点研发计划项目(2024YFC3211600) （2024YFC3211600）
中国地质调查局黑河流域水循环野外站联合开放基金项目(WCSHR-2024-08) （WCSHR-2024-08）
