Next Article in Journal
Winter Wheat Yield Prediction at County Level and Uncertainty Analysis in Main Wheat-Producing Regions of China with Deep Learning Approaches
Previous Article in Journal
Verification of Fractional Vegetation Coverage and NDVI of Desert Vegetation via UAVRS Technology
 
 
Letter
Peer-Review Record

Automatic Gully Detection: Neural Networks and Computer Vision

Remote Sens. 2020, 12(11), 1743; https://doi.org/10.3390/rs12111743
by Artur M. Gafurov * and Oleg P. Yermolayev
Reviewer 1:
Reviewer 2:
Reviewer 3: Anonymous
Remote Sens. 2020, 12(11), 1743; https://doi.org/10.3390/rs12111743
Submission received: 22 April 2020 / Revised: 23 May 2020 / Accepted: 25 May 2020 / Published: 28 May 2020
(This article belongs to the Section Environmental Remote Sensing)

Round 1

Reviewer 1 Report

This manuscript describes the use of ANN to develop a model to interpret aerial photographs for the presence of gullies. It reports some success in achieving these aims.

 

The authors do an excellent job of explaining the context of the research and the necessity of developing a new technique, given the lack of a high resolution DEM for Russia.

The study relies on training data created by experts (there's really no way around this that I see), but the researchers indicate that the experts appear to have missed identifying some existing gullies, resulting in positive identifiers by the ANN, but negative by the experts (training data). These appear as false negatives, but are simply missed by the experts. 

The ANN also appears to identify some linear features as gullies when they are in fact not gullies (false positive). These are improvements that will be required of the ANN before it can be usefully applied to other regions, however, the authors acknowledge this and propose a solution (additional training runs with new training data).

One of the conclusions presented is that rills and gullies are properly distinguished by the ANN model (i.e., the model ouput does not mis-identify rills as gullies). In the methods and results, I was unable to find where this was discussed, it just appeared in the conclusions. I feel this must be better supported within prior sections of the manuscript.

Throughout the manuscript, terms were presented without definition nor explanation. For example, 

  • gully network dynamics in Figure 1 (also, what are polygons in Figure 1?) The explanation of the negative values for dynamics was unclear
  • dissection density (line 141)
  • loamy loam (line 162)
  • power (line 162)
  • sum of biologically active temperatures (line 169) - is it as defined in Grigorieva et al. 2010? found here https://www.int-res.com/articles/cr2010/42/c042p143.pdf) If so, then biologically active temperatures depends on a particular crop. Perhaps reporting mean annual and summer/winter temperatures would be more appropriate. 
  • continental climate coefficient (line 169)
  • flood flow module (line 173)
  • annual flow module (line 174)
  • gully etalon (line 212) is this a typo? This word exists, but apparently not within the context of gully erosion.

Most of the bullet points listed above are also marked on the attached PDF. 

Several other comments, questions, and corrections are also marked on the pdf.

Overall I think the research described in this manuscript is useful and will be of interest. It presents a novel and potentially groundbreaking application of ANN for gully detection, especially in areas without a high resolution DEM that could make use of existing automated gully detection methods.

Comments for author File: Comments.pdf

Author Response

Thank you very much for the review of the article and the important corrections to the text. We tried to take into account all the comments.

Throughout the manuscript, terms were presented without definition nor explanation. For example, gully network dynamics in Figure 1 (also, what are polygons in Figure 1?) The explanation of the negative values for dynamics was unclear dissection density (line 141)

To unify the terms used, we have adopted the term "dissection density" throughout the text, which is the total length of the entire gully network in the river basin, divided by its area. In Figure 1, the polygons represent river basins (we have made adjustments in the figure), negative values mean a decrease in the gully density of gully dissection compared to the previous time interval. This term is the most commonly used.

loamy loam (line 162)

Technical error: there are loams

power (line 162)

We agree with the comment, that term has been replaced by a more general one (thickness).

The sum of biologically active temperatures (line 169) - is it as defined in Grigorieva et al. 2010? found here https://www.int-res.com/articles/cr2010/42/c042p143.pdf) If so, then biologically active temperatures depend on a particular crop. Perhaps reporting mean annual and summer/winter temperatures would be more appropriate.

Agreed. This indicator is used as a standard for the physical and geographical characteristics of territories in Russia. Indeed, it would be better to give air temperature characteristics instead.

In the text added: the average annual air temperature -3.9 ° C, the average temperature in January -11.9 ° C, July +19.1 °, the average annual long-term amplitude of air temperatures 53.8 °.

continental climate coefficient (line 169)

At its calculation the coefficient proposed by Yu. G. Simonov (1972) was used, which combines temperature conditions and atmospheric humidification of the cold period of the year.

Kк = (mк+nк); mк=(1-tj/to)*cosY; nк=(ro(год.)/riх.п.)-2,

where tj and to is the average January temperature at a given point and in the open ocean of the temperate zone (to = +5); ro(год.) = 1000 mm is the annual precipitation in the open ocean; riх.п. - precipitation for the cold period of the year at a given point; Y - latitude.

An explanation has been added to the text: «The continental climate coefficient that includes both temperature conditions and atmospheric humidification for the cold period of the year is 2.1 - 2.2.»

flood flow module (line 173)

In our opinion, it is not reasonable to give interpretation in the text, because the meaning of this standard hydrological indicator is embedded in the units of measurement of the parameter.

In other words, it is the volume of water during the period of meltdown flowing from a certain basin area to a unit of time. It is measured in cubic meters per second from a square kilometre (m³/(s × km²), and for small values measured in liters per second from a square kilometre - l/(s × km²). It is calculated by dividing the water runoff by catchment area using the following formula (in m³/(s × km²)): M=Q/F, where Q is water consumption m³/s; and F is catchment area, km².

annual flow module (line 174)

The indicator calculates the annual water runoff according to the same formula, but only for a year. The runoff coefficient is the ratio of the size of a layer of runoff from a given area for a certain period of time to the size of a layer of atmospheric precipitation falling on the same area for the same period of time.

All these indicators relate to the standard parameters of hydrology and cannot raise any questions from the readers of "Remote Sens.

gully etalon (line 212) is this a typo? This word exists, but apparently not in the context of gully erosion.

No, it's not a typo. The gully etalon samples have indeed been prepared. Such reference objects (samples) were visually identified by experts on satellite images of gullies. Many of such forms were verified in the field and therefore they can be considered as reference objects and used for neural network training.

Other improvements:

Figures 1-3 are moved over the text and renamed as recommended

Figure 2 - the legend has a section explaining what polygons (river basins) mean

Line 173-174 "flood flow module" corrected to "river runoff unit"

Line 184 by "decade" we mean 10 days. We agree - replaced by "half."

Figure 7 - Since the reviewer had questions about the content of the map and considering that improvement of GECNN is ongoing, changes were made to the figure. It becomes clear from the map why we claim that GECNN perfectly differentiates between gullies and rills, as all the erosive forms typical of a given area can be seen in the figure. In particular, the rills that flow into the gullies were not marked out, unlike the latter.

Reviewer 2 Report

The aim of this study was to develop a methodology for automated detection of gully erosion based on remote sensing data using neural networks.

The aim is clearly stated at the last paragraph of introduction but then it was also mentioned the 6 methodological steps of the research that should be part of the methods chapter and not as aim of the research. The aim should be mentioning also what could be the practical implications of this novel methodology.

The state of the art was briefly described and the current practice was mentioned showing the gap that will be the aim of the research.

The method was adequately described.

Figure 2 is not related to the aim of this study and can be deleted.

The discussion of the results was insufficient. It was not mentioned any comparison with other studies on gully erosion and it was not shown the advantages of the proposed methodology in relation to other methods and in different context. Is the method adequate for agricultural land or for forest land?

It seems that all test area was low slope but if it could be tested in medium and high slope then maybe it will be found significant differences in gully detection. If the authors have available data set from higher slopes then they should check if there are differences in accuracy of gully detection between low slope and high slope areas.

It seems that the test area is uniform land use (agriculture ploughed land). If the authors have available data set from different land use or different soil management practice at the study area then they may search for differences in gully erosion created due to wrong land management practices. Then this new method can be used for early detection of gully erosion due to wrong soil management and propose changes that will protect soil. Is the method good to detect gullies in different land use and land management practices? 

Author Response

Figure 2 is not related to the purpose of this study and can be deleted.

We would like to keep in the article Fig.2. This is, first of all, unique historical data on land use change. Secondly, it indirectly shows the level of anthropogenic (primarily agricultural) load on landscapes. It was the deforestation after the land reform in Russia in 1861 that led to uncontrolled ploughing of territories (steep slopes) and appearance of most modern gullies. Also visually well seen is the relationship between the dynamics of the gully network and changes in land use (forestry to agriculture).

The discussion of the results was insufficient. It was not mentioned any comparison with other studies on gully erosion and it was not shown the advantages of the proposed methodology in relation to other methods and in different contexts. Is the method adequate for agricultural land or for forest land?

On lines 37-61 there was an analysis of gully erosion studies using other methods in the European part of Russia, similar in landscape conditions to the test area. In the section "Discussion" a detailed comparative description of the results obtained using GECNN and GEOBIA(OBIA) for the same area was given. To summarize, OBIA is not able to analyze abstractions, which is important when scaling the classifier to other areas, when GECNN uses patterns and graphs. If we consider forest areas separately, if the gully is not visible in the image by an expert, it will not be recognized by the neural network. In the case of forested areas, we need a DEM, which, under conditions of frequent planting, can only be obtained using total stations or GNSS.

It seems that all test areas were low slope but if it could be tested in medium and high slope then maybe it will be found significant in gully detection. If the authors have available data set from higher slopes then they should check if there are differences in accuracy of gully detection between low slope and high slope areas.

In answer to this question, let us refer to a quote added in the section "Discussion": “Assessing the prospects of GECNN scaling, as well as the possibility of applying the neural network in other territories, it should be noted that in the current implementation of the algorithm is aimed at identifying gullies in the Republic of Tatarstan, which is representative of the European part of Russia. Considering that images of gullies representing abstract forms are analyzed, there is no difference for the neural network, what type of soil is depicted and what humidity conditions are in the study area (within the European part of Russia). However, in case of necessity, additional training of the model with the use of images of gully development conditions specific for a particular territory will allow, in our opinion, to extend GECNN application possibilities”. The territory of the Republic of Tatarstan in the area chosen for testing does have low slopes (median ~2 degrees, Q3 = 2.69 degrees). The territory of the EPR has a median basin slope of 1.19 degrees, Q3 = 1.8 degrees). However, it is wrong to assume that gullies develop on plain parts of slopes. The slope in the gullies varies from 10 to 25 degrees, and the variety of samples used to train the neural network, allows to identify any gullies that have certain interpretation features - the vertex, a clearly defined boundary, visible talveg.

It seems that the test area is uniform land use (agriculture ploughed land). If the authors have available data set from different land uses or different soil management practices at the study area then they may search for differences in gully erosion created due to wrong land management practices. Then this new method can be used for early detection of gully erosion due to poor soil management and propose changes that will protect soil. Is the method good to detect gullies in different land uses and land management practices?

The dominant type of land use in the test region is agricultural (farming). In addition, gullies resulting from this type of land use are the dominant part of all gullies. The share of technogenic gullies (mainly formed by the road network) is less than 1%. Technogenic gullies are single cases of erosive forms, but they have high growth rates (linear and area growth). At the same time, morphologically they do not differ from agricultural type gullies and therefore the proposed method can be used for identification in various land use and land management practices.

Reviewer 3 Report

This paper is original and meets aims and scope of the journal. Authors use clear and unambiguous, professional English throughout the draft. Also, they provide literature references, sufficient field background/context. I think the structure of the article conform to an acceptable format as well.

Transition from visual interpretation to fully automated gully detection is an important task for quantitative assessment of modern gully erosion, especially when it comes to large mapping areas. However, application of neural networks for object detection and contouring on satellite images is poorly developed so far. This research could provide a methodology, which is of significance, and seems very interesting and relevant in soil erosion and other purposes. At this moment, the effort of authors is appreciated but still needs minor improvement as described below prior to acceptance.

 

  1. This is a case study, is it possible to apply the study results to other places worldwide? Not only limited to the territory of the European part of Russia.
  2. In the present study, ultra-high resolution satellite image data were selected, so more precise and clear the images were used, more accuracy the results can be obtained, but only focused on the small scale, how to apply them to the larger scale? This is important.
  3. In the study area, wind erosion is apparently not significance, it seems that both water and tillage erosion have great impact on this area, and these two erosion forms interact and then cause more soil erosion. My question is which erosion form is more important (water or tillage) in this area?
  4. If tillage operation happens first, then comes the rainfall event, which generates the gully, and vice versa. Can the training results of GECNN be affected by such a circumstances? In other words, selected images based on time and location is quite important, for they will definitely have impact on the training results.
  5. Is running code the open source (Python)? As the authors say problems still exist, and it is necessary to additionally train the neural network on the enlarged training data set.

Author Response

Thank you very much for your careful review of the manuscript.

Answering the questions, I would like to note at once that the answers to them formed the basis for additions in the " Discussion" section.

This is a case study, is it possible to apply the study results to other places worldwide? Not only limited to the territory of the European part of Russia.

So, in answer to the first question, here's a quote: «Assessing the prospects of GECNN scaling, as well as the possibility of applying the neural network in other territories, it should be noted that in the current implementation of the algorithm is aimed at identifying gullies in the Republic of Tatarstan, which is representative of the European part of Russia. Considering that images of gullies representing abstract forms are analyzed, there is no difference for the neural network, what type of soil is depicted and what humidity conditions are in the study area (within the European part of Russia). However, in case of necessity, additional training of the model with the use of images of gully development conditions specific for a particular territory will allow, in our opinion, to extend GECNN application possibilities».

In the present study, ultra-high resolution satellite image data were selected, so more precise and clear the images were used, more accuracy the results can be obtained, but only focused on the small scale, how to apply them to the larger scale? This is important.

Answering the second question, in addition to the above, it should be noted that we used RGB-synthesis of DigitalGlobe images with 60 cm resolution. This cannot be called a small scale, because the resolution, nadir directionality of the image and high-precision positioning of the image bring them closer to orthophotoplans. We agree that it is possible to use larger resolution images, but we have a number of reasons why we chose 60cm resolution. First, since we use 512*512 pixels fragments as samples, the 60 cm resolution of the image allows us to display a significant part of the gully in the fragment. Secondly, increasing the image resolution multiplies the required computational power, which in our case is limited. Thirdly, the available experience of expert gully recognition on satellite images allows concluding that this resolution is more than enough for accurate gully boundaries recognition.

In the study area, wind erosion is apparently not significance, it seems that both water and tillage erosion have great impact on this area, and these two erosion forms interact and then cause more soil erosion. My question is which erosion form is more important (water or tillage) in this area?

Unfortunately, there is no easy way to answer this question. Historically, anthropogenic activities in this area contributed to the development of gully erosion processes. Nowadays, when anthropogenic loads are stabilized and in conditions of changing climate, water component gradually takes the first role.

If tillage operation happens first, then comes the rainfall event, which generates the gully, and vice versa. Can the training results of GECNN be affected by such a circumstances? In other words, selected images based on time and location is quite important, for they will definitely have impact on the training results.

In answer to this question, let us repeat the above quotation from the extended section "Discussion": «Assessing the prospects of GECNN scaling, as well as the possibility of applying the neural network in other territories, it should be noted that in the current implementation of the algorithm is aimed at identifying gullies in the Republic of Tatarstan, which is representative of the European part of Russia. Considering that images of gullies representing abstract forms are analyzed, there is no difference for the neural network, what type of soil is depicted and what humidity conditions are in the study area (within the European part of Russia). However, in case of necessity, additional training of the model with the use of images of gully development conditions specific for a particular territory will allow, in our opinion, to extend GECNN application possibilities». The variety of gullies used for training makes it possible to assert that it does not matter what caused the gully. All gullies are described by the specific interpretation features - the presence of the vertex, a clearly defined boundary, talveg. These very abstractions are analyzed by GECNN.

Is running code the open source (Python)? As the authors say problems still exist, and it is necessary to additionally train the neural network on the enlarged training data set.

We are considering releasing the written code to the open source as soon as the neural network training and gully recognition work in the Republic of Tatarstan is completed.

Round 2

Reviewer 1 Report

I appreciate that the authors highlighted the changed text in the revised manuscript. This helped to reduce the review time. 

The additional discussion is also helpful.

The revised manuscript has addressed nearly all of the concerns and comments from the first review. I have made note of a few outstanding comments in the attached pdf using comments, but these are very minor and can be addressed at the editor level.

I recommend accepting after these minor revisions/corrections are addressed.

Comments for author File: Comments.pdf

Author Response

Line 58. Added "For example" to explain the types of unresolved problems in office mapping of gullies.

Line 171. River runoff, as a standard hydrological variable, is the volume of water during the period of meltdown flowing from a certain basin area to a unit of time. It is measured in cubic meters per second from a square kilometre (m³/(s × km²), and for small values measured in liters per second from a square kilometre - l/(s × km²). It is calculated by dividing the water runoff by catchment area using the following formula (in m³/(s × km²)): M=Q/F, where Q is water consumption m³/s; and F is catchment area, km².

Line 229. Added missing a closing bracket.

Reviewer 2 Report

The paper was significantly improved and can be accepted in the present form.

Author Response

Thank you very much for reviewing the manuscript and helping to improve it.

Back to TopTop