Improving crop classification with landscape stratification based on MODIS-time series
Stratification is needed to accurately classify crop types in heterogeneous landscapes, but required data for stratification is limited for many regions. Freely available MODIS imagery may fill this void. In this study, stratification methods based on unsupervised classification of MODIS time-series imagery were compared to stratification methods based on detailed soil- and elevation maps. To this end, stratified classification accuracies of WorldView 2 & 3 imagery were compared for a series of twelve images covering the case study area of Sougoumba, Mali.
A per pixel EVI vegetation index value was derived for every fortnight MODIS composite covering the years 2001-2015. Three stratification layers were constructed based on the average- , the amplitude- , and the seasonality respectively. The first stratification layer was built by simply taking a per pixel mean of the EVI images throughout this time period. The second stratification layer was calculated using a BFAST analysis, in which a time series is decomposed into a trend, season and remainder. From this, the seasonal component was used to calculate the amplitude (Figure 1). The last stratification layer was built by comparing the slope in greenness of a three- time step moving window to the slope in greenness from winter to summer. The first window that has a steeper slope than the winter- to summer slope was taken as the start of the growing season (Figure 1).
Figure 1. On the left you can see the calculation of the amplitude from BFAST. On the right you can see the estimation of the length of the growing season.
Subsequently, the three maps were clustered into three distinct categories using k-means clustering. The resulting maps are shown in Figure 2.
mean (based on the average of the EVI time series)
amplitude (based on the difference between the peak and lowest point of the growing season in terms of the EVI).
seasonal (based on the length of the growing season).
Figure 2. The three stratification maps after k-means clustering of EVI time-series into three clusters.
Afterwards, WorldView images were classified on a training set of 3792 samples and validated using a training set of 1881 samples.
It was clear that stratification led to an increase in classification accuracy, and also that the amplitude- and length-of-growing-season-method performed best (Figure 3).
Figure 3a. The different classification methods and their corresponding overall accuracy. This figure is the result of a Random Forest classification of the image of the 29th of July, 2014.
Figure 3b. A comparison of the overall accuracy of a single image (29-07-2014) vs. a multi-temporal NDVI composite of 2014. The stratification method used for these calculations is Amplitude, and the classification algorithm is Random Forest.
In this study, single multi-band image classification using the Random Forest classifier increased from 56% without stratification to 61%, 67% and 65% using the average, amplitude and length-of-growing-season stratifications respectively, which is slightly better than the results obtained using soil and elevation strata in a previous study by Ommeren (2016).
Classification accuracies of NDVI composite time-series using with the Random Forest classifier resulted in lower overall classification accuracies when compared to the PVI Ommeren (2016). However, the increase in accuracy from non-stratified to stratified classification was similar in both studies.
In conclusion, MODIS time series can be a very useful tool for developing stratification layers, and provide a viable alternative to stratification techniques based on local soil maps. The stratification methods proposed in this study yielded satisfying classification accuracies and have a large potential for upscaling.
Still, interpreting strata derived from MODIS time-series remains challenging and requires input from local experts with knowledge of the landscape, especially when moving towards zones with multiple growing seasons per year.