Mapping of multitemporal rice (Oryza sativa L.) growth stages using remote sensing with multi-sensor and machine learning : a thesis dissertation presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy in Earth Science at Massey University, Manawatū, New Zealand

Figure 2.1 is adapted and re-used under a Creative Commons Attribution 4.0 International (CC BY 4.0) license.Rice (Oryza Sativa) plays a pivotal role in food security for Asian countries, especially in Indonesia. Due to the increasing pressure of environmental changes, such as land use and climate, rice cultivation areas need to be monitored regularly and spatially to ensure sustainable rice production. Moreover, timely information of rice growth stages (RGS) can lead to more efficient of inputs distribution from water, seed, fertilizer, and pesticide. One of the efficient solutions for regularly mapping the rice crop is using Earth observation satellites. Moreover, the increasing availability of open access satellite images such as Landsat-8, Sentinel-1, and Sentinel-2 provides ample opportunities to map continuous and high-resolution rice growth stages with greater accuracy. The majority of the literature has focused on mapping rice area, cropping patterns and relied mainly on the phenology of vegetation. However, the mapping process of RGS was difficult to assess the accuracy, time-consuming, and depended on only one sensor. In this work, we discuss the use of machine learning algorithms (MLA) for mapping paddy RGS with multiple remote sensing data in near-real-time. The study area was Java Island, which is the primary rice producer in Indonesia. This study has investigated: (1) the mapping of RGS using Landsat-8 imagery and different MLAs, and their rigorous performance was evaluated by conducting a multitemporal analysis; (2) the temporal consistency of predicting RGS using Sentinel-2, MOD13Q1, and Sentinel-1 data; (3) evaluating the correlation of local statistics data and paddy RGS using Sentinel-2, PROBA-V, and Sentinel-1 with MLAs. The ground truth datasets were collected from multi-year web camera data (2014-2016) and three months of the field campaign in different regions of Java (2018). The study considered the RGS in the analysis to be vegetative, reproductive, ripening, bare land, and flooding, and MLAs such as support vector machines (SVMs), random forest (RF), and artificial neural network (ANN) were used. The temporal consistency matrix was used to compare the classification maps within three sensor datasets (Landsat-8 OLI, Sentinel-2, and Sentinel-2, MOD13Q1, Sentinel-1) and in four periods (5, 10, 15, 16 days). Moreover, the result of the RGS map was also compared with monthly data from local statistics within each sub-district using cross-correlation analysis. The result from the analysis shows that SVM with a radial base function outperformed the RF and ANN and proved to be a robust method for small-size datasets (< 1,000 points). Compared to Sentinel-2, Landsat-8 OLI gives less accuracy due to the lack of a red-edge band and larger pixel size (30 x 30 m). Integration of Sentinel-2, MOD13Q1, and Sentinel-1 improved the classification performance and increased the temporal availability of cloud-free maps. The integration of PROBA-V and Sentinel-1 improved the classification accuracy from the Landsat-8 result, consistent with the monthly rice planting area statistics at the sub-district level. The western area of Java has the highest accuracy and consistency since the cropping pattern only relied on rice cultivation. In contrast, less accuracy was noticed in the eastern area because of upland rice cultivation due to limited irrigation facilities and mixed cropping. In addition, the cultivation of shallots to the north of Nganjuk Regency interferes with the model predictions because the cultivation of shallots resembles the vegetative phase due to the water banks. One future research idea is the auto-detection of the cropping index in the complex landscape to be able to use it for mapping RGS on a global scale. Detection of the rice area and RGS using Google Earth Engine (GEE) can be an action plan to disseminate the information quickly on a planetary scale. Our results show that the multitemporal Sentinel-1 combined with RF can detect rice areas with high accuracy (>91%). Similarly, accurate RGS maps can be detected by integrating multiple remote sensing (Sentinel-2, Landsat-8 OLI, and MOD13Q1) data with acceptable accuracy (76.4%), with high temporal frequency and lower cloud interference (every 16 days). Overall, this study shows that remote sensing combined with the machine learning methodology can deliver information on RGS in a timely fashion, which is easy to scale up and consistent both in time and space and matches the local statistics. This thesis is also in line with the existing rice monitoring projects such as Crop Monitor, Crop Watch, AMIS, and Sen4Agri to support disseminating information over a large area. To sum up, the proposed workflow and detailed map provide a more accurate method and information in near real-time for stakeholders, such as governmental agencies against the existing mapping method. This method can be introduced to provide accurate information to rice farmers promptly with sufficient inputs such as irrigation, seeds, and fertilisers for ensuring national food security from the shifting planting time due to climate change