National farm scale estimates of grass yield from satellite remote sensing

Loading...
Thumbnail Image
Files
116220859-Richa Marwaha.pdf(10.64 MB)
Full Text E-thesis
Date
2021-12-03
Authors
Marwaha, Richa
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Research Projects
Organizational Units
Journal Issue
Abstract
Globally, grasslands are an important source of food for livestock and provide additional ecosystem services such as greenhouse gas (GHG) mitigation through carbon sequestration, habitats for biodiversity, and recreational amenities. Grass is the cheapest source of fodder providing Irish farmers with an economic benefit against international competitors. Hence, to maintain profitability, farmers have to maximize the proportion of grazed grass in cow’s diet or save it as silage. The overall objective of the current research project was to build a machine-learning model to estimate grass growth nationally using earth observation imagery from the Sentinel 2 satellite constellation and ancillary meteorological data, which are known to influence grass growth. Firstly, the impact of meteorological data and Growing Degree Days (GDD) was assessed for Teagasc Moorepark experimental farm (Fermoy, Co Cork, Ireland). GDD was modified to include Soil Moisture Deficit (SMD), which included the impact of summer drought conditions in 2018. Results demonstrated the importance of GDD for grass growth estimation using ordinary linear regression (OLS). The potential evapotranspiration (PE) 0.65 (r=0.65) and evaporation (r=0.65) were equally significant variables in 2017, while in 2018 the solar radiation had the highest correlation (r=0.43), followed by potential evapotranspiration and evaporation with r of 0.42. The standard and modified GDD were equally significant variables with r of 0.65 in 2017, but both had a reduced correlation in 2018 with modified GDD (0.38, p<0.01) performing slightly better than the standard GDD (0.26, p<0.01) calculation. These models only explained 53% (RMSE of 18.90 kg DM ha-1day-1) and 36% (RMSE of 27.02 kg DM ha-1day-1) of variability in grass growth for 2017 and 2018, respectively. Considering the importance of meteorological data, an empirical grass model called the Brereton model, previously used for Irish grass growing conditions were tested. Since this model lacks a spatial element, we compared the Brereton model with the previously used machine-learning model ANFIS and Random Forest (RF) with the combination of satellite data and meteorological data for eight Teagasc farms. Overall, the machine-learning algorithms (R2= 0.32 to 0.73 and RMSE=14.65 to 24.76 kg DM ha-1day-1 for the test data) performed better than the Brereton model (range of R2=0.03 to 0.33 and RMSE=41.68 to 82.29 kg DM ha-1day-1). The RF model (with all the variables except rainfall) had the highest accuracy for predicting grass growth rate, with (R2= 0.55, RMSE = 14.65 kg DM ha-1day-1, MSE= 214.79 kg DM ha-1day-1 versus ANFIS with R2 = 0.47, RMSE = 15.95 kg DM ha-1day-1, MSE= 254.40 kg DM ha-1day-1). When developing a national model, meteorological data were missing (except precipitation). A different approach was followed, whereby the grass growing season was subdivided (January-June Agmodel 1 and July–December Agmodel 2). Phenologically, the peak grass growth in Ireland typically occurs in May, with a slow decline in subsequent months. Spring is the most important season for grassland management, where growing conditions can impact the grass supply for the whole year. The national models were developed using Sentinel 2 band metrics, spectral indices (NDVI and NDRE), and rainfall for 179 farms. Data from 2017-2019 was divided into training and testing data (70:30 split), with 2020 data used for independent validation of the final trained model. Test accuracy was higher for Agmodel 1 (R2 = 0.74, RMSE= 15.52 kg DM ha-1day-1) versus Agmodel 2 (R2 = 0.58, RMSE= 13.74 kg DM ha-1day-1). This trained model was used on validation data from 2020, and the results were similar with better performance for Agmodel1 (R2 =0.70) versus Agmodel2 (R2=0.36). The improved spatial resolution of Sentinel 2 and the availability of red-edge bands showed improved results compared with previous work based on coarse resolution satellite imagery.
Description
Keywords
Remote sensing , Grassland , Machine-learning , Sentinel-2 , Grass growth
Citation
Marwaha, R. 2021. National farm scale estimates of grass yield from satellite remote sensing. PhD Thesis, University College Cork.