RETRACTED: A Participatory Iterative Mapping Approach and Evaluation of Three Machine Learning Algorithms for Accurate Mapping of Cropping Patterns in a Complex Agro-Ecosystems

Abstract

We applied time-series analysis of vegetation indices (VIs) (NDVI and EVI) derived from the Moderate Resolution Imaging Spectrometer (MODIS) sensors to detect seasonal patterns of irrigated and rainfed cropping patterns in five townships in the central dry zone of Myanmar, an important agricultural region of the country, that is both poorly mapped for cropping practices and which faces environmental and climate related challenges to agriculture. To improve mapping accuracy of cropping pattern, we implemented a participatory iterative ground truthing and mapping approach and explored the efficiency of three state-of-the-art machine learning algorithms: Support Vector Machine (SVM), Random Forest (RF), and a classification tree and rule-based model (C5.0). We first collected reference data at random locations and run a preliminary supervised classification using the SVM algorithm. Based on the preliminary classification outputs, we invited township agricultural officers to assess the accuracies of the maps based on local knowledge and secondary statistical data and hence identify areas with high land cover heterogeneity, which enabled us to allocate more sample sizes in such areas. We compared accuracies achieved by use of increasing size of predictor layers of VIs (8-days, 16-days and monthly composite stacks of 1 to 3 years). Results show the combined effects of i) an iterative participatory approach to field data collection and map classification, ii) identification of superior algorithm and iii) appropriate size and type of predictor VIs, we were able to substantially improve mapping accuracy; depending on the models used, accuracy improvement ranged from 31% to 43%. Among the three algorithms we compared, SVM with Gaussian Radial Basis kernel function was superior in terms of all accuracy measures. Particularly, the accuracy difference was statistically significant (p < 0.005) when larger numbers of VIs layers were used and significance of difference diminishes with the increasing size of training data. Accuracy achieved by use of NDVI was consistently better than that of EVI. Though the difference is not significant, 8-days NDVI composite resulted in better accuracy than 16-days composite. Maximum overall accuracy of 94% was achieved using SVM and 8-days NDVI composites of three years. In conclusion, our findings highlight the value of participatory field validation approaches to improve classification accuracy especially in areas where land use patterns are temporally and spatially complex. We also show that choice of classifiers and size of predictor variables are essential and complementary to the participatory field approach in achieving desired accuracy of cropping patter mapping in areas where other sources of spatial information are scarce.