Just a few days after publishing my latest paper I came across a similar paper that had been published recently:
“Sociopolitical influences on cropland area change in Iraq, 2001–2012” by Gibson, Campbell and Zipper.
This article investigates changes in cropland area in Iraq and their connections to the sociopolitical situation in the country during this period. They find that cropland were generally decreasing by as much as 30 000 ha per year. In the Kurdish region, cropland decreased by about 16 500 ha/year. This decrease in cropland is explained by the increasing imports to the area, which have reduced the need for domestic production.
Now, I’m not focusing on the whole Iraq or the whole Kurdistan Region so it might be a scale issue, but if you have skimmed my most recent publication on cropland you might have noticed that I measured a potential increase in cropland between the 98-02 and the 11-14 periods.
Why are our results contradicting?
Gibson et al. (2015) uses the MODIS Land Cover Type product (Short Name: MCD12Q1) that is a ready-to-use product holding data on land cover. The product has a spatial resolution of 500 m and provides data with a yearly frequency between 2001 and 2012. Five global land cover classification systems are included, and classifications are based on the spectral and temporal information from the 7 bands, the vegetation indices and the land surface temperature. In Gibson et al. (2015), the Plant Functional Type (PFT) classification scheme was used because it distinguishes between broadleaf crops and cereal crops, and was found to be largely agreeing with the International Geosphere-Biosphere Programme
(IGBP) classification scheme and the University of Maryland (UMD) scheme.
Leroux et al. (2014) assessed the reliability of the MODIS LCP in sub-Saharan Africa by comparing the product to agricultural statistics data (FAOSTAT: & AGRHYMET) and high resolution satellite images (SPOT & Landsat). They find that MODIS LCP’s ability to estimate cropland area is good at a regional and national scale levels, but at more detailed levels it should be used with caution.
I became curious about how well MODIS LCP performed in my study area, an area that I know quite well on the ground and from satellite images. Below is an image from Google Earth of the Duhok Governorate.
I downloaded data from the USGS website (https://lpdaac.usgs.gov/data_access/data_pool) for the years 2001, 2002, 2011 and 2012. These years corresponded to two of the periods I had focused on in my recent article, period C (1998-2002) and period D (2011-2014).
In the maps above I have overlayed the two years to show pixels classified as cereal cropland during period C and D. I don’t know about you but to me it looks like someone has taken a giant lawn mower and ran it over all cropland in the south western part of the Governorate some time before 2011. Now look at the satellite image again. Can you see any rectangle pattern characteristic for cropland? Where? Now take a look at my classification (below).
This classification is based on Landsat which has a spatial resolution of 30 m. It has an overall accuracy of between 93.4 and 97.1 % (depending on which period), and a Kappa value of between 60 and 86%. Gibson et al. (2015) mention that the MODIS LCP data might only include large scale croplands, but unfortunately it looks like it finds these areas in mountain areas, where I know for sure that no such large croplands exist. It might be a good idea to also consult the GlobeLand30 dataset, which I wrote about in a post last year, that corresponded reasonably well with my ground truth points. The strong decrease that Gibson et al. find in the LCP images is definitely visible, but I’m afraid that in the case of the Duhok governorate, it is probably more due to classification errors than sociopolitical factors.
Gibson et al. present the results of an accuracy assessment of the LCP, based on visual inspection of high resolution satellite imagery – similar to the validation I did in my paper. However, they only present the overall accuracy, which is between 92 and 96 percent (just slitghtly lower than my overall accuracy, presented above). What I think should be clarified here is that the overall accuracy does not say much about the classification scheme’s ability to classify cropland correctly, just that the overall classification, the average classification accuracy was high. This means that the classification of non-cropland might have a very high accuracy, whereas the cropland accuracy might be lower. User’s and producer’s accuracies should have been presented to give a better idea of the quality of the classification, and I wonder why it wasn’t. To me it is not reasonable to have an accuracy of 95% for a global classification.
So what do we get out of this comparison? I think Gibson et al. has made an interesting and important study, and at national level the trends detected give some idea of the changes to cropland area during the past decade. However, like Leroux et al. (2014) found, at a more detailed level the data used might not be reliable, at least not in the Duhok governorate (or sub-Saharan Africa), and therefore a more detailed presentation of the accuracy assessment would be desirable.
This example also highlights the relevance of scale in geographical research, and how the overall, large-scale, spatial and temporal patterns might not be representative at a more detailed level. It shows how important it is to be critical about secondary data and how it was produced. This is the topic of my last paper in my thesis that will hopefully be published soon.