the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Apr 2024
Technical Note: A simple feedforward artificial neural network for high temporal resolution classification of wet and dry periods using signal attenuation from commercial microwave links
Abstract. Two simple feedforward neural networks (MLPs) are trained to classify wet and dry periods using signal attenuation from commercial microwave links (CMLs) as predictors and high temporal resolution reference data as target. MLPGA is trained against nearby rain gauges and MLPRA is trained against gauge-adjusted weather radar. Both MLPs perform better than existing methods, showcasing their effectiveness in capturing the intermittent behaviour of rainfall. This study is the first using both radar and rain gauges for training and testing for CML wet-dry classification. Where previous studies has mainly focused on hourly reference data, our findings show that it is possible to predict wet and dry periods with a higher temporal precision.
- Preprint
(621 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2024-647', Anonymous Referee #1, 02 May 2024
Publisher’s note: a supplement was added to this comment on 7 May 2024.
See the reviewer comments in the attached
- AC1: 'Reply on RC1', Erlend Øydvin, 27 Jun 2024
-
RC2: 'Comment on egusphere-2024-647', Anonymous Referee #2, 23 May 2024
Comments to Technical Note: A simple feedforward artificial neural network for high temporal resolution classification of wet and dry periods using signal attenuation from commercial microwave links. by Oydvin et al. (2024)
Application of neural network technic on the attenuated signals of commercial microwave links was introduced to estimate the non-precipitation periods during the intermittent rainfall observed by the gauge based measurements. Data sets with 3901 CMLs for a month were analyzed (trained) with references of one minuets interval rain gauge data and high-temporal 5-minuets interval gauge-adjusted radar products, and statistical differences and temporal characteristics of two products (MPL-GA and MPL-RA) were revealed. Estimating of high-temporal precipitation variabilities using the CMLs might be quite promising technique to produce remote area’s precipitation fields with difficulties of ground observation or variation of satellite products at near surface level. MLP would be helpful to translate the CMLs attenuation signals into the duration/amount of precipitation quantitatively. This technical note introduced the performance of MLP comparing with former methods proposed by Graf et al. (2020) and Polz (2020). However, I could not catch up quantitively how better the proposed methods were from the contents. Why is estimating the intermitted precipitation important and which kinds of products are expected to be produced by the MPLs? There are many uncertain technical descriptions to understand the data process. Therefore, the note needs fundamental revision before the formal publication on the HESS.
General comments
1) “wet and dry periods” is hard to understand. There must be a miss-conversion of concepts between “spatial no-precipitation events identified by CML” and “temporal no-precipitation records by the gauge”. In the CMLs, “a dry period” may corresponds to a period without precipitation along one link. However, the authors convert the term to the “dry periods” as multiple no-precipitation periods in a time sequence of point-measured intermittent data. There are no detailed explanation of how the 3901 links data were integrated and aggregated by the MLP? Besides, a gauge measurement could only provide the precipitation intensity over the site without any spatial information of the rain system. “precipitation and non-precipitation periods” in the time sequence are virtual signals, and they are depending on the sensor types and data recording interval (I could not understand “minuets interval 0.01 mm resolution gauge”, did you use disdrometer or tipping bucket event recorder?). For instance, non-precipitation periods change depending on one-minuets, hourly, daily, or monthly data even in the same location. Please clarify the concepts of “wet and dry periods” to be understood in the sense of meteorology/hydrology with corrections of title and usage in the contents.
2) I could not understand what do you want to “classify”? Do you want to distinguish rain or no-rain of instantaneous CML signal, or you want to adjust the CML signals to the precipitation intensity? Also, what does “feedforward” mean? Do you want to propose the new methods or just demonstrate how the MLP works? Please clarify the target (objects) of this technical note.
3) It is concluded that MLPs performed “better” than existing methods. Some statistics showed differences between your products and sigma80/CNN products, but comparisons on long-term sequence failed to estimate the variations (L184, Fig. 5). Please show the clear reasons of why your work is “better”, and explain them in the conclusion. You propose the benefit of higher temporal estimates (L6-7), however it looks like failing filter out the short-term noise in Fig.5. Also, if you proud of the higher temporal product shorter than one hour, it is better to scale up the Figure 3 with one-minuets interval.
Specific comments
Title: Not clear the meaning of “feedforward”, “wet and dry period”,”
L9 “signal attenuation” of what?
L17 more and more “available” mean data are becoming open?
L19 What is the “wet period”, “Based on this,”?
L20 You frequently use “prediction”, but this term is for the future estimation. Better to reconsider the usage as “estimate” such as in L175.
L32 “hourly reference data” means “hourly precipitation data”? Is this gauge data? No study using disdrometer?
L38 What is “rainy time step”? Non-liquid precipitation means solid precipitation such as snow fall?
L40-41 I could not understand the meaning.
L43 Object of this note is to describe the methods? Please clarify the object here according to the conclusion.
2.1 Data: Please clarify the detailed network structure of the CML, such as general distance, distribution, and what is the “near the CML”? Which kind of rain gauge is used? How the instruments could get 1-minuet/0.01mm resolution? Is this optical sensor? 5 minuets resolution of DWD stans for every 5 minuet interval or 5 minuet average?
L60 “wet” may mean the existence of precipitation record. Then, there must be two kind of “wet periods” such as gauge based and radar based? Both periods were defined by the same threshold as 0.01mm per a minuet?
2.3-2.5 Some of this part are separated only by the paragraph. Please reconsider to combine the sub-section into one section composed by the story only by paragraphs.
L81 As the comparison of your methods to two previous method is the key, please explain in detail about two reference methods (Sigma80 and CNN).
L112-119 This part describe the study direction, and better to move before. “intermittent rainfall” is your focus, so better to explain what it is (how did you detect).
L122 Delete “given optimal”. Such type existed in may places.
L126 What is “spatial difference”?L131 You insist “more consistently” from which part of the figure? Need more polite explanation to the readers.
L150- I could not understand the process making Fig. 3-5. Are they case studies in different time series, or some kind of composite with different time scale? Better to divide upper and lower part as Fig. 3a and 3b, but the lower parts are not explained fully in the contents. White areas are dry periods?
L155 Why “nicely”? I can not catch up which part of the figure could correspond to your results. Same asn L161 “Further ,,, rise”, etc.
L161 “Further, ,,rise.” I could not identify those tendency in Fig. 3. Need marks or specify with times.
Fig.3 upper: Clear characteristics on Fig.3, that I could identify, were such as 1) CNN is similar to GA and Sigma80 is similar to RA, 2) MPL-GA and MPL-RA estimated longer no-precipitation periods as RA, 3) MPL-GA and MPL-RA looks mostly similar with a small difference around 6:00. If the MLP studied the RA and GA observations respectively, why the MLP-GA and MLP-RA are so similar? Your focuses may be on more small scale, but I could not figure out the importance of your argument. The graph of GA should put above the RA to adjust the order of MPL-GA and MPL-RA.
L168 I could not identify the wet starting point.
L170 I could not understand the description of “non of the methods ,,”. What is the “reference period”?
L173 “However, “ I could not understand what you mean.
L178-182 This part should move to the paragraph in L166.
Fig.5 This is the result of longest record with few precipitation periods in RA GA where MPL reproduced some intermittent periods of rains similar to sigma80. The author mentioned “erratic signals” in L184, and attributed by a noisy CML signal. It means the MLP even could not filter out one noise on the CML? Then why you can conclude “better (L3)” performance?
L190-195 Future issues should be mentioned in the conclusion.
Citation: https://doi.org/10.5194/egusphere-2024-647-RC2 - AC2: 'Reply on RC2', Erlend Øydvin, 27 Jun 2024
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 241 | 58 | 20 | 319 | 13 | 15 |
- HTML: 241
- PDF: 58
- XML: 20
- Total: 319
- BibTeX: 13
- EndNote: 15
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1