University of Twente Student Theses


Reconstruction of vegetation indices using multi-source images and deep learning - A case of study in the Netherlands

Hidalgo Sulca, S.A. (2023) Reconstruction of vegetation indices using multi-source images and deep learning - A case of study in the Netherlands.

[img] PDF
Abstract:The projected increase in the global population by 2050, coupled with the impacts of climate change, presents a significant challenge of meeting global food demand while minimizing the risk of hunger. To achieve sustainable development and address the second Sustainable Development Goal of zero hunger, advancements in technology, policy, and governance are crucial. The European Union's Common Agricultural Policy (CAP) introduced the "greening" initiative in 2013 to promote sustainable agricultural productivity. However, the effectiveness of these measures in improving biodiversity and ecosystem services has been limited due to various challenges, including heterogeneous management rules, lack of spatial targeting, and costly field inspections. To overcome these challenges, the utilization of Earth Observation data, particularly from the Sentinel-1 and Sentinel-2 satellites, has emerged as a promising solution. For example, the Sen4CAP project aims to develop algorithms and products based on Sentinel-1 and Sentinel-2 time series data to monitor agricultural activities at the parcel level, thereby supporting the greening policy. While vegetation indices (VIs), such as the normalized difference vegetation index (NDVI), derived from Sentinel-2 data, are commonly used to analyze vegetation properties, cloud cover poses a significant obstacle. Conversely, the use of Sentinel-1 (SAR) data provides advantages such as weather independence but requires interpretation due to signal complexity. Combining the strengths of both sensors has shown promising results in tasks such as mowing detection, crop monitoring, and crop mapping, although challenges persist. Deep learning approaches, including generative adversarial networks (GANs), have gained popularity in remote sensing applications by demonstrating their potential in reconstructing missing data. A Multi-Temporal Conditional Generative Adversarial Network (MTcGAN) approach that combines SAR and optical data has been proposed to reconstruct VIs. This method considers two sensing times, utilizing SAR data from t1/t2 and optical data at t1 to simulate optical data at t2. Additionally, studies employing GANs for image reconstruction have shown improved performance across different crop types. The experiments conducted in this study involve the reconstruction of VIs using optical and SAR data in the Flevoland province in the Netherlands. Experiment A focuses on reconstructing optical bands at different times and subsequent VIs, while Experiment B aims to directly reconstruct VIs using VIs and SAR data. These experiments are compared with similar research conducted in the field. The evaluation of results is performed using metrics such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), along with visual evaluation metrics. In Experiment A, the RGB bands yielded higher evaluation metric values compared to the NIR band, subsequently affecting the calculation of VIs derived from them. Visual evaluation of RGB compositions demonstrated the model's potential to generate data under cloud cover in unseen data, albeit with certain limitations and mispredictions. Experiment B, conducted with a smaller training sample, exhibited unsatisfactory performance. Notably, variations in the size of the training dataset, sensing interval between t1 and t2, and pre-processing techniques were observed among reference studies. Reducing the sensing interval has the potential to enhance performance while using data from different locations but within a close sensing time could also yield improved results. Although the study area lacked cloud-free images during the more developed phenology stages, using data from partially clouded images remains a viable option, with the extent of feasibility yet to be determined. Consequently, the VIs did not exhibit significant differences between crops, with NIR predictions still proving to be the least accurate. However, patches of cropland exhibited fewer mispredictions, indicating that contextual information about crops within predominantly cropland areas enhances prediction accuracy.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Subject:48 agricultural science
Programme:Spatial Engineering MSc (60962)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page