Earth observations for official crop statistics in the context of scarcity of in-situ data

Abstract

Remote sensing offers a scalable and low cost solution for the production of large-scale crop maps, which can be used to extract relevant crop statistics. However, despite considerable advances in the new generation of satellite sensors and the advent of cloud computing, the use of remote sensing for the production of accurate crop maps and statistics remain dependant on the availability of ground truth data. Such data are necessary for the training of supervised classification algorithms and for the validation of the results. Unfortunately, in-situ data of adequate quality for producing crop statistics are seldom available in many countries. In this paper we compare the performance of two supervised classifiers, the Random Forest (RF) and the Dynamic Time Warping (DTW), the former being a data intensive algorithm and the latter a more data frugal one, in extracting accurate crop type maps from EO and in-situ data. The two classifiers are trained several times using datasets which contain in turn an increasing number in-situ samples gathered in the Kashkadarya region of Uzbekistan in 2018. We finally compare the accuracy of the maps produced by the RF and the DTW classifiers with respect to the different number of training data used. Results show that when using only 5 and 10 training samples per each crop class, the DTW reaches a higher Overall Accuracy than the RF. Only when using five times more training samples, the RF starts to perform slightly better that the DTW. We conclude that the DTW can be used to map crop types using EO data in countries where limited in/situ data are available. We also highlight the critical importance in the choice of the location of the in-situ data and its thematic reliability for the accuracy of the final map, especially when using the DTW.