Classification and Mapping of Paddy Rice using Multi-temporal Landsat Data with a Deep Semantic Segmentation Model

As a staple commodity in major world markets, paddy rice plays an important role in various environmental issues including food security, water use, climate change, and disease transmission. Therefore, timely and accurate classification and monitoring on paddy rice are necessary for scientific and practical purposes. Precise paddy rice field maps require high spatial resolution and temporal observation frequency. Moderate Resolution Imaging Spectroradiometer (MODIS) images at 500-m or 250-m spatial resolution have problems with coarse spatial resolution and mixed land cover types within a pixel. When generating annual cropping maps at large scales using multi-temporal and high spatial resolution (e.g. Landsat TM/ETM+), the problems with low temporal resolution and missing pixels(e.g. cloud cover and SLC-off data) reduce the amount of usable data. In the past, the accuracy and efficiency for paddy rice classification and mapping at fine spatial resolutions were limited by the poor data availability and traditional classification methods. Thanks to the strong ability of deep learning(DL) in feature representation, DL has been introduced into environmental remote sensing and applied in many aspects, including land cover mapping. This study aims to use a semantic segmentation model for large scale paddy rice classification based on multi-temporal Landsat data. To address the challenge of Landsat data due to low temporal resolution and inevitable cloud contamination, this research uses linear interpolation and temporal resampling to largely overcome the cloud contamination issue. The experiment was carried out in Arkansas located in the middle and lower reaches of the Mississippi River, with an area of 137,733 km2, which is the largest rice producing state in the U.S. Experiments were designed to evaluate what information is most useful for training the deep learning model for paddy rice classification, and how various temporal intervals (7-day, 14-day and 28-day) affect the paddy rice classification performance in order to derive optimal time interval. All experiments were conducted over Arkansas’ agricultural planting areas, and a total of 63 Landsat multitemporal scenes including all the six optical spectral bands spanning from 2016 to 2019 were used to train and test. To assess the performance of the trained model, we calculated pixel-based metrics and analysed the results by using overall accuracy (OA) and F1 score. Computational experiments show temporal resampling leads to improved classification accuracies when compared to non-resample. A good performance of the proposed model was also recorded when temporal interval was 7-day and 14-day. Compared with 6 optical spectral bands as training features, 3 bands (red, near-infrared and short-wave infrared) show notably better performance and have overall accuracy of 0.77-0.91, with the F1 score of 0.65-0.91. Overall, our results demonstrate the valuable potential of the semantic segmentation model that utilizes multi-temporal data from Landsat which allows for detailed classification and mapping of paddy rice field over large areas.