Multi-Temporal Landsat-8 Images for Retrieval and Broad Scale Mapping of Soil Copper Concentration Using Empirical Models

Round 1

Reviewer 1 Report

In this study, the authors proposed a framework for retrieving and mapping soil copper concentration based multi-temporal RS images. The manuscript is interesting, and contains description of new approach of analysis of soil heavy metal assessment. However, there are some issues that need to be resolved before this manuscript can be considered for publication. My recommendation is MAJOR revision.
General comments:

1. As REOTE SENSING is an international journal, it seems to be better to give soil names in accordance with international soil classification (WRB).
2. The language needs to be edited by native speakers. The use of language should be carefully checked. There are many unnecessary descriptions. There are more deficiencies in language use. English polishing is must.
3. In general, the logics among all sections are poor. Therefore, I can hardly get your objectives in this review.
4. The manuscript contains many bad-quality figures. For example, the similar contents in Figs. 2 and 3, and the strange design of Fig.8. For Figs.4 and 5, Bar charts are preferable.
5. The abstract lacks the necessary structure, usually the abstract includes background, purpose, methods, results and conclusions/research implications. In fact, I can not catch any effective information here.
6. The uncertainty is another key issue in soil mapping, and the simulation with a low uncertainty is needed. The uncertainty is always evaluated by repeatedly conducting the simulations. In this study, the authors adopted machine-learning method, which may generate high uncertainty, and thus, the uncertainty evaluation is necessary. However, the authors even did not evaluate the uncertainty. The result generated by the simulation only at one time was adopted as the final result, which is inappropriate and considerably influences the reliability of the results.
7. Generally, There should be a consideration of the seasonal variation in soil Cu magnitude and its spectral signature on images.
8. For 2. Theoretical background, it is crucial. However, Cu is not a spectrally active component. The authors should make most efforts to clarify the estimating mechanism. In addition, you should also prove why the introduction of multi-temporal RS images improve the inversion. These problems are crucial in your study.

Specific comments:

1. The introduction needs to be expanded.
2. Please provide a bit more big-picture motivation of how your analyses benefit society and how they have evolved over the past decade. However, from my point of view, the article does not provide a sufficiently thorough review of the issue under study. There are good references for the study techniques, but the paper is missing a "big-picture" introduction with some references in my opinion. I suggest that the authors should do a better analysis of the literature. It seems that the bulk of the text is a sort of compilation of statements in the individual articles cited. It would be better, I think, to extract ideas from individual articles and tie them together into a more fluid and conceptually homogeneous text. As it is, the text looks rather clumsy.
3. Research gaps, objectives of the proposed work should be clearly justified before the problem formulation section. This paper includes some little useful information and the main objectives of the study is not well defined. Problem statement is not clear and the objectives are obscure. Furthermore, the paper lacks a very clear and good justification for what is new and innovative about this case or this approach.
4. For soil heavy metal inversion and uncertainty analysis, some recent articles available for similar studies can be read.
   a) Assessing toxic metal chromium in the soil in coal mining areas via proximal sensing: Prerequisites for land rehabilitation and sustainable development[J]. Geoderma, 2022, 405: 115399.
   b) Digital mapping of zinc in urban topsoil using multisource geospatial data and random forest[J]. Science of The Total Environment, 2021, 792: 148455.
5. c) Ensemble machine-learning-based framework for estimating total nitrogen concentration in water using drone-borne hyperspectral imagery of emergent plants: A case study in an arid oasis, NW China[J]. Environmental Pollution, 2020, 266: 115412.
6. Explain how soil Cu is determined and provide laboratory error.
7. For quantifying study, only R2, RMSE, ME are not sufficient. Please use RPD or RPIQ. In fact, these two are more effective measures in remote sensing inversion.
8. The illustration of kriging and related semi-variance function should be included in your study. Because they are very related to your interpolation in Fig. 8.
9. The detailed information soil sampling campaign and experimental design should be added.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Manuscript Recommendation

Review of Remote Sensing manuscript 1677319 – “Multi-temporal Landsat-8 images for the retrieval and mapping of the copper concentration in soil using empirical models”.

 

This research evaluates an agronomically, ecologically, and environmentally important soil property -- heavy metal content (HMC) or concentration using remote sensing imagery and spectral observations (Landsat-8). The authors employ sequenced approach that employs multi-temporal satellite observations to improve spectral information followed by statistical approaches to evaluate prediction features and identify and evaluate the optimal regressor for soil HMC prediction from spectral data in the area studied.  The research approach is scientifically valid, and results receive appropriate and adequate statistical development and evaluation. The manuscript lacks a few key details that are important for data interpretation and complete understanding of the methods and study results. Detailed review comments follow that outline needed provide details on needed additions and and/or edits. The manuscript is appropriate for publication in Remote Sensing following revision. The manuscript in places has awkwardly written English that and contains extra verbiage. The abstract is rephrased as an example to help clarify the writing.

 

 

Detailed review comments.

 

Title and subsequent text

The term “copper concentration in soil” is long and oft repeated in the manuscript.

The term can be shortened to soil copper concentration.

As an example, the title can be rephrased -- Multi-temporal Landsat-8 images for retrieval and broad scale mapping of soil copper concentration using empirical models.

 

Rephrased Abstract

Retrieval and mapping of soil heavy metal concentration from satellite remote sensing images using empirical models is a crucial and challenging task. Such retrieval relies on efficient data preparation, feature extraction and selection, as well as development of an optimal empirical model. To establish a reliable inverse model for soil copper (Cu) concentration estimation using regression methods, this paper proposes a retrieval and mapping framework that uses Landsat multi-temporal images and integrates the following strategies. First, we employ multi-temporal satellite observations to enhance spectral information of images. Second, key prediction features are extracted using principal component analysis (PCA), minimum noise fraction transform (MNF) and isometric feature mapping (ISOMAP) algorithms, and effective features are selected as the model input. Third, we compare multiple regressors (i.e., support vector regression (SVR), partial least square regression (PLSR) and artificial neural network (ANN)) unbiasedly to obtain the optimal regressor for estimating the soil Cu concentration within the study area. Data used in this study include 11 Landsat-8 images from 2013 to 2017 in Gulin County, Sichuan China, and 138 soil samples collected from the area in 2015 that have lab measured Cu concentrations. Study results indicate that compared to a single image (SI) of Landsat, use of several temporal images increases the estimation accuracy from 0.433 to 0.641, as measured by the mean adjusted coefficient of determination (R²) obtained by SVR using 20 repeated 6-fold cross-validation. Due to the feature extraction using PCA, MNF and ISOMAP algorithms, as well as feature selection based on permutation feature importance, the mean R²of PLSR and ANN increase respectively from 0.568 to 0.618 and from 0.476 to 0.528. Using the cross-validation estimation, regression models are compared unbiasedly based on their best performances with their favorite feature combinations. The best regressor--SVR achieves a mean R²of 0.641, outperforming the worst regressor, i.e., ANN, by 21.4% percentage. Other measures, e.g., the root mean square error (RMSE), mean absolute error (MAE) and standard error (SE), also improve. A preferred approach that generated this highest regression success involves SVR model trained with original bands of temporal images. The estimated soil Cu content in the study area shows a spatial pattern that is consistent with the land cover classification map. Our results indicate that the combined improvements in data preparation, feature extraction and classifier selection can increase the prediction accuracy of empirical models, and as such the proposed framework has a great potential for supporting large-scale operational retrieving of soil heavy metal concentration from Landsat-8 images.

 

Introduction

Page 2-3

Lines 87-92

 

The text in lines 87-92 discuss study results and should be omitted from the introduction.

 

Study Area and Soil Samples.

 

Figure 2. This figure presents the study area size and location within the province. An additional inset is needed to show the province location within China.

 

Additional soil sample information is needed. Specifically, the sample depth, quantity collected, sample and prep, (air-dried? 2mm sieving?).

 

The sampling points are not uniformly distributed. What rationale was used to select the sample points? Ease of access? Land use? Other logistics?

 

Metals in soils are measured by various techniques and chemical extractions that have different intent and outcomes. The specific method employed for the Cu determination in this study needs more complete description and a cited reference provided. This is a key detail because the subsequent model development and prediction equations are developed from measured sample values.

A table providing descriptive statistics (mean, median, std dev, mode, maximum, minimum) of the measured soil data should be included in the manuscript.

 

 

Page 10

Figure 7 - The axes for predicted and measured should be labelled with the units (e.g., mg/kg).

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The revision was sufficiently revised according to my previous comments. I think it is acceptable for publication in Remote sensing.