Abstract. Drought is a highly consequential natural disaster that may likely increase in both severity and extent across the conterminous United States (CONUS). The mechanisms affecting the propagation of drought from the atmosphere to streamflow are complex and interactive, making the prediction of streamflow drought difficult with current modeling approaches. Machine learning is an emerging tool in the field of hydrology that may be well-suited to prediction of streamflow drought across large and topographically diverse areas. Here, we train and analyze 3,198 random forest models at U.S. Geological Survey streamgages to understand common meteorological drivers of streamflow drought and to define physiographic characteristics of basins sensitive to these drivers. We also develop a novel dynamic regionalization approach using donor gages to predict daily streamflow drought at pseudo-ungaged locations. For this study, CONUS was divided into nine regions for ease of reference in describing results. Our results show that teleconnections, temperature, evaporative demand, and snow-water equivalent are important drivers of streamflow drought in the West, Southwest, and Northern Rocky Mountains (Northern Rockies) regions of the United States, and precipitation and soil moisture are primary drivers of streamflow drought in the Northeast, Southeast, and the Northwest regions. Prediction using dynamic regionalization shows comparable performance to at-site models.
Received: 04 Dec 2025 – Discussion started: 21 Jan 2026
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
This work analyzes and predicts streamflow drought in the CONUS via machine learning and statistics. Random forests are trained to predict droughts from meteorological forcings. Gini impurity values are computed to investigate variable importance at the regional scale. Principal component analysis (PCA) on the variable importances allows the investigation of large-scale drivers of drought across the CONUS. Linear regression is used to predict principal components (PCs) from basin characteristics in ungaged basins. Similarity between PCs is used to select donor gages from which forcings are translated to the ungaged basin and converted by the random forest into drought prediction.
The study is well executed and very well documented. The motivation and objectives are clear from the beginning. The methodology is explained in detail and likely allows for reproduction. The dicussion nicely picks up on the research objectives, offers explanations for surprising findings, and draws elaborate connections to existing literature.
Specific comments (most of them are suggestions and therefore optional): - l 23-24: Sentence about dividing CONUS into nine regions could be removed - l 61-62: Indeed, ML requires large amounts of data processing during training. Once a model is trained, though, costs are often amortized - l 62-63: Sentence about interpretability feels out of place here, though interpretability is important, so I would put it somewhere else - l 66-68: Should conflicting training signals from different basins be called noise? Also, isn't the main advantage of the donor-based approach that it allows PUB? - l 81: Optionally replace "faced across the nation" with "present across the country" or similar - l 88: Did you mean non-uniform? - ll 87-91: I feel a little lost with this information and I don't think it is picked up again. Are you trying to say that the results should be taken with a grain of salt? - ll 105-107: Once you understand that sentence, it is clear, but it took me a while. Maybe the formula could be extended a bit to make it easier - l 131-132: Comment: Also, if you would include antecedent streamflow, the task would be much easier since droughts are defined based on streamflow, and the model likely wouldn't learn the connections you're after - Table 1: I would add horizontal lines to delimit different sources and references. Also, it would be nice if the table wasn't split across two pages - ll 150-152: What exactly is meant with robustness of the method, especially considering that robustness to redundancies and correlations are separately mentioned right after? - ll 156-162: This could be shortened a bit - l 169: Did you play around with the node size parameter and found this value empirically or is this a common choice? - l 180 and l 176 are the same formula - ll 191-192: But didn't you explicitly reweight the data so that the classes are balanced? Most likely I'm missing something here, so maybe you could elaborate a bit - ll 198-201: It was not directly obvious to me how the enumeration relates to the figure. I think adding the paragraph numbers to the enumeration sentence would already help a lot - Figure 2: This figure contains more than the "development and evaluation of dynamic regionalization method", which is the name of this subsection. It contains all the study's research questions! So I would suggest to reference it not only in that subsection but more generally. Also, the figure is almost completely linear, so I think simply plotting it from top to bottom would be more intuitive than the current layout, but that might be a personal preference - ll 248-250: These sentences seem redundant to me - l 256: Similarly to node size, I'm wondering whether the 1% were found empirically - l 259: I understand that you found 0.35 empirically here, but wouldn't one expect it to be 0.2 given you drought definition? On the other hand, you reweighed your training data, so one might even expect a threshold of 0.5. Interestingly, 0.35 is right in between. Is there something to it, or did I miss something? - Figure 4: First (and only) occurrence of "Julian Day" is in this caption, maybe introduce it earlier. Also, how does it relate to "Decimal Date"? - ll 313-314: You explain what loadings are here, but you already used the term in l 229 - Figure 7: If I understand correctly, sometimes, prediction works better with the donor approach than with an at-site models. I think this should be discussed, what could be the reason for this? - ll 546-550: The conclusion is mostly on a rather high level (which is good), so I wouldn't explicitly mention PC1 etc. here, because those terms are quite technical. Maybe there is a way to phrase it more generally
Technical corrections: - l 293: "Decimal Date" -> "Decimal date" - l 350: Remove closing parenthesis - l 409: "Date" -> "date" - l 596: doi should be "https://doi.org/10.1175/1520-0493(1987)115<1083:CSAPOL>2.0.CO;2" - l 603: doi should be "https://doi.org/10.1175/1520-0493(1969)097<0163:ATFTEP>2.3.CO;2" - l 635: doi should be "https://doi.org/10.1175/1520-0477(1998)079<2715:IMOET>2.0.CO;2"
All in all, this is a very good study in my opinion. After some minor revisions, which mostly concern intelligibility, it is ready for publication in HESS.
We used machine learning to explore what causes streamflow droughts across the U.S. We found that different regions are influenced by different factors like temperature, snow, and rainfall. Our new method can also predict droughts in areas without streamflow data, helping improve water resource planning.
We used machine learning to explore what causes streamflow droughts across the U.S. We found...
This work analyzes and predicts streamflow drought in the CONUS via machine learning and statistics. Random forests are trained to predict droughts from meteorological forcings. Gini impurity values are computed to investigate variable importance at the regional scale. Principal component analysis (PCA) on the variable importances allows the investigation of large-scale drivers of drought across the CONUS. Linear regression is used to predict principal components (PCs) from basin characteristics in ungaged basins. Similarity between PCs is used to select donor gages from which forcings are translated to the ungaged basin and converted by the random forest into drought prediction.
The study is well executed and very well documented. The motivation and objectives are clear from the beginning. The methodology is explained in detail and likely allows for reproduction. The dicussion nicely picks up on the research objectives, offers explanations for surprising findings, and draws elaborate connections to existing literature.
Specific comments (most of them are suggestions and therefore optional):
- l 23-24: Sentence about dividing CONUS into nine regions could be removed
- l 61-62: Indeed, ML requires large amounts of data processing during training. Once a model is trained, though, costs are often amortized
- l 62-63: Sentence about interpretability feels out of place here, though interpretability is important, so I would put it somewhere else
- l 66-68: Should conflicting training signals from different basins be called noise? Also, isn't the main advantage of the donor-based approach that it allows PUB?
- l 81: Optionally replace "faced across the nation" with "present across the country" or similar
- l 88: Did you mean non-uniform?
- ll 87-91: I feel a little lost with this information and I don't think it is picked up again. Are you trying to say that the results should be taken with a grain of salt?
- ll 105-107: Once you understand that sentence, it is clear, but it took me a while. Maybe the formula could be extended a bit to make it easier
- l 131-132: Comment: Also, if you would include antecedent streamflow, the task would be much easier since droughts are defined based on streamflow, and the model likely wouldn't learn the connections you're after
- Table 1: I would add horizontal lines to delimit different sources and references. Also, it would be nice if the table wasn't split across two pages
- ll 150-152: What exactly is meant with robustness of the method, especially considering that robustness to redundancies and correlations are separately mentioned right after?
- ll 156-162: This could be shortened a bit
- l 169: Did you play around with the node size parameter and found this value empirically or is this a common choice?
- l 180 and l 176 are the same formula
- ll 191-192: But didn't you explicitly reweight the data so that the classes are balanced? Most likely I'm missing something here, so maybe you could elaborate a bit
- ll 198-201: It was not directly obvious to me how the enumeration relates to the figure. I think adding the paragraph numbers to the enumeration sentence would already help a lot
- Figure 2: This figure contains more than the "development and evaluation of dynamic regionalization method", which is the name of this subsection. It contains all the study's research questions! So I would suggest to reference it not only in that subsection but more generally. Also, the figure is almost completely linear, so I think simply plotting it from top to bottom would be more intuitive than the current layout, but that might be a personal preference
- ll 248-250: These sentences seem redundant to me
- l 256: Similarly to node size, I'm wondering whether the 1% were found empirically
- l 259: I understand that you found 0.35 empirically here, but wouldn't one expect it to be 0.2 given you drought definition? On the other hand, you reweighed your training data, so one might even expect a threshold of 0.5. Interestingly, 0.35 is right in between. Is there something to it, or did I miss something?
- Figure 4: First (and only) occurrence of "Julian Day" is in this caption, maybe introduce it earlier. Also, how does it relate to "Decimal Date"?
- ll 313-314: You explain what loadings are here, but you already used the term in l 229
- Figure 7: If I understand correctly, sometimes, prediction works better with the donor approach than with an at-site models. I think this should be discussed, what could be the reason for this?
- ll 546-550: The conclusion is mostly on a rather high level (which is good), so I wouldn't explicitly mention PC1 etc. here, because those terms are quite technical. Maybe there is a way to phrase it more generally
Technical corrections:
- l 293: "Decimal Date" -> "Decimal date"
- l 350: Remove closing parenthesis
- l 409: "Date" -> "date"
- l 596: doi should be "https://doi.org/10.1175/1520-0493(1987)115<1083:CSAPOL>2.0.CO;2"
- l 603: doi should be "https://doi.org/10.1175/1520-0493(1969)097<0163:ATFTEP>2.3.CO;2"
- l 635: doi should be "https://doi.org/10.1175/1520-0477(1998)079<2715:IMOET>2.0.CO;2"
All in all, this is a very good study in my opinion. After some minor revisions, which mostly concern intelligibility, it is ready for publication in HESS.