Early yield estimation by photosynthetic pigment abundances using landsat 8 image series

Date

2020-10-13

Author

Özcan, Ayşenur

Metadata

Show full item record

Item Usage Stats

49
views

28
downloads

Cite This

Timely estimation of crop yields is critical for monitoring global food production by international organizations as well as governments, farmers and the private sector dealing with storage, import and export of crops and associated products. Satellite remote sensing has the capability to provide near real-time information on a global scale. Combining satellite data and soft computing techniques to predict crop yields is a very effective strategy for continually forecasting crop yields. This thesis presents a novel approach for accurate and sustainable estimation of crop yields based on estimated abundances of endmembers that may be attributed to photosynthetic pigments. Landsat 8 images acquired during the time of the phenological cycle when plants have maximum greenness are the inputs to find endmembers and abundances within the pure wheat crop pixels using Robust Collaborative Nonnegative Matrix Factorization (R-CoNMF) unmixing algorithm. The endmembers are optimized to maximize the predictive power of the abundances for the yields. Wheat yields were then estimated with the four abundances, their relevant interactions, ten important agrometeorological vi parameters, including parameters proposed in this thesis for the first time, and four different vegetation indices using three different machine learning algorithms (Generalized Linear Model (GLM), Artificial Neural Network (ANN) and Random Forest (RF)). Harvester records from 142 wheat fields distributed in 31 provinces of Turkey were used as the ground truth for testing the algorithm. In the literature, the coefficient of determination (R2 ) is used as a proxy to show how good the relationship is between the estimated and real figures. According to these calculations, the yields were estimated with 64% accuracy when only the abundances were used in the GLM algorithm, 78% accuracy when ANN was used for yield estimation and 82% accuracy was reached when applying RF to all of the parameters. The similarity of the endmembers to photosynthetic pigment spectral signatures along with their predictive power suggested their relevance to the pigments. Although the R-CoNMF algorithm performs a linear unmixing of the intimate mixture of the photosynthetic pigments, the interactions of the abundances used in the endmember optimization and in classifications partially handle the nonlinearity using the bilinear model. These results can be considered as a great success when using multispectral satellite data only and are recognized as a clear indication that much better results would be achieved while using images from future hyperspectral space missions like HyspIRI.

Subject Keywords

Landsat 8, Time series, Yield estimation, Random forest, Artificial neural network, Generalized linear model, Photosynthetic pigments, Unmixing, RCoNMF, Endmember optimization, Endmember extraction

URI

https://hdl.handle.net/11511/69144

Collections

Graduate School of Natural and Applied Sciences, Thesis