Past and Future Trajectories of Farmland Loss Due to Rapid Urbanization Using Landsat Imagery and the Markov-CA Model: A Case Study of Delhi, India

Round 1

Reviewer 1 Report

This is an interesting paper based on the analysis of Landsat images to predict trends in farmland loss in Delhi. This research area has considerable technical interest but also the implementation of reliable models can be of major significance for government planning. The paper describes a novel approach to the analysis of the satellite imagery, validates the model and uses it to predict future trends.

The overall structure of the paper is good and it contains a clear introduction to the field and also to the specific techniques being used in the modelling. A very useful and comprehensive list of references is provided. The paper is readable, requiring comparatively minor copy editing (it would be good if a native English reader could glance through the paper and make the minor changes, but the meaning is clear even without such changes). There is good use of diagrams and the results are presented well. The conclusions are well-considered.

I have a very few minor points. On line 158, I'm not sure what the 15m error refers to. Is this an issue of image registration? And on line 359, what does "dominant change" mean? Is this actual area or proportion of area? In glancing through the references (which I didn't really check, there's a typo on line 592.

Overall I'm happy with this paper - it describes a nice piece of work.

Author Response

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents the past and future trajectories of farmland loss in Delhi, India using the Markov-CA model under multiple constraints. The topic is interesting and the experiments presented in this paper are logical. However, there are some major points required to be addressed.

1.     Compared to previous widely used Markov-CA methods, what’s the advantage of your study? Please highlight this point in your experimental tests and results.

2.     Lines 88-90. It seems that incorporating socioeconomic or demographic variables into simulation models will be promising. However, I can find no details emphasis on this. In the meantime, the comparison between predicted maps A and B (Figure 7) also shows the inconsistency of performance improvement. What’s the major drivers behind? It is worth more explanations and discussion.

3.     The paper uses MLC to obtain LULC maps for the model development. The calibration and validation steps are all based on these classified maps. Thus, the classification accuracy and consistency will be fundamentally important. Have you checked the inconsistency of post-classification between different years? This issue is in particular, critical to LULC change detection and prediction. How do you tackle with this issue?

4.     In Table 2, the accuracy assessment of classified maps indicates a relatively low accuracy for farmland and grassland. Does it matter to your model prediction?

5.     Lines 448-459, Page 12: From Table 6, we can see that the "full" model is slightly better than the "only" model. It's hard to judge whether the improvements result from the full model or from the relatively low classification accuracy of the validation test. Therefore, I strongly recommend the authors to analyze and discuss this issue in the discussion.

6.     Lines 511- 513, Page 13: Farmland loss has been attributed to rapid urbanization in this section. However, from Table 3 and 4, major types transited from "Farmland" are "Urban" and "Forest" (322.35 v.s. 200.77, 391.12 v.s. 236.98). From this perspective, the contribution to farmland loss is not fully explained and should be further discussed.

7.     Figure 6. The prediction performance is not so plausible from the error maps, especially for 2014. It seems that the prediction error mainly occurs within grassland and farmland. I wonder it may be due to the low classification accuracy of these two land cover types.

8.     Lines 314-318. The accuracy assessment of model performance should be revised. The RMSE derived from the percentage for each class will underestimate the prediction biases. You should provide more spatial explicit approach to account for the difference, and also calculate the pixel-based prediction biases for each class.

9.     For Remote Sensing journal, the geographic coordination is required for maps and remote sensing imagery.

Author Response

Author Response File: Author Response.pdf