Ashish Vivekar (2020) Evaluation of methodology for estimating crop yield from multispectral UAV images : a case study at Lönnstorp, Sweden

Sammanfattning: Adaptation of modern Unmanned Aerial Vehicle (UAV) and multispectral sensor technology in agriculture can enhance the capacity to accurately monitor crops. But these technologies come with its own set of challenges. The major challenge is the understanding of radiometric distortions, which is particularly important while comparing data over different lighting conditions. The study developed and assessed a methodology for extracting radiometrically corrected reflectance values for multi-temporal datasets and to find out if crop yield it can be estimated using it. The empirical line method is used to calibrate the images using spectrally stable panels. The methodology is assessed by studying the association between three vegetation indices namely Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI) and Difference Vegetation Index (DVI) with corresponding crop yield, crop height and fixed tower sensor data. The radiometric correction techniques delivered reasonably satisfactory results which was revealed by very strong Pearson correlation (r = 0.87 - 0.95) between fixed tower and UAV sensors. The investigation also identified that the vegetation indices dependency varied positively with (fresh) yield of legume ley, from moderate to high, on the first harvest date (11th May 2019) as well as on the second harvest date (26th July 2019). The Pearson correlation (r) and Spearman rank correlation coefficients (rs) ranged between 0.55 and 0.89 on the first harvest date and between 0.49 and 0.79 on the second harvest. Comparable results were obtained for other crops but only at certain stages of crop development. The analysis revealed moderate to high positive relationship (r = 0.45 – 0.87) between vegetation indices (NDVI and SAVI) and crop height except on 21st August 2019 dataset, where the relationship was rather weak (r = 0.19 and 0.04). DVI showed similar trend (r = 0.62 – 0.81), except in the case of 26th July 2019 and 21st August 2019 datasets, where correlation coefficients were r = 0.19 and r = -0.02 respectively. It was also observed that 6 datasets over the growing period are not enough to clearly see the complete crop phenology, although the comparison of NDVI, SAVI and DVI indicated similar patterns. The inability in reflecting the complete phenology of various crops was mainly due to less frequent data acquisition at various development stages. Nonetheless, very high positive correlation between vegetation indices from UAV sensor and fixed tower sensor validated the capability of UAV sensor to monitor the crops over various stages. Keywords: Physical Geography and Ecosystem analysis, UAV, Multispectral Sensor, Radiometric Correction, NDVI, SAVI, DVI, Phenology. Advisor: Lars Eklundh Master degree project 30 credits in Geo-information Science and Earth Observation for Environmental Modelling and Management (GEM), 2019 Department of Physical Geography and Ecosystem Science, Lund University. Student thesis series INES nr 28