Rice-planted area mapping using ALOS-2 PALSAR-2 data with machine learning in Southeast Asia

Abstract:

Rice is a staple cereal crop in Asia, and the continent accounts for about 90% of global rice production and consumption. Rice-planted area map is important parameter to estimate rice production for food security or economy, and also to quantify the carbon, water cycle or methane emission via paddy fields. Rice is mainly cultivated in the rainy season, Synthetic Aperture Radar (SAR) is therefore a robust tool because it penetrates cloud cover. Recently, machine learning has been widely used in many land cover related researches and distinct results were reported, however, limitation is that it needs a large amount of training data, it is normally time and cost consuming task. In this research, we utilized the combination of unsupervised and supervised classification to efficiently produce the training data. Training data were generated from the k-means classification results for the sampled regions, then a random forest classifier was applied to ALOS-2 PALSAR-2 ScanSAR data to identify rice-planted area in Southeast Asian countries. It is also difficult to identify rice-planted area in this region since there are high variations in rice phenology. In order to compensate for the variations, we used time-series metrics of calculated from SAR data. Classification models were fine-tuned for each country, and most of the models had an accuracy of 0.9 or better. Independent verification through visual interpretation using very high resolution images (VHRs) on Google Earth also showed a high degree of consistency with the classification results. The developed paddy field maps showed high accuracy in most countries and regions, however, verification using in-situ data and national statistics, as well as application to other regions, seasons and years, is necessary to confirm the effectiveness of proposed methodology. It would be also promising methodology to improve classification accuracy by using the latest machine learning algorithms or by combining C-band SAR data and optical satellite data with ALOS-2.