Remote sensing and machine learning for prediction of wheat growth in precision agriculture applications

  • Yuxi Fang

Student thesis: Master's Thesis

Abstract

This thesis focuses on remote sensing and machine learning for prediction of wheat growth in precision agriculture applications.Agriculture is the primary productive force, which plays an important role in human activities. Wheat, as one of the essential sources of food, is also a widely planted crop. The impact of weather and climate and some other uncertain factors on wheat production is crucial. Therefore, it is necessary to use reliable and statistically reasonable models for crop growth and yield prediction based on vegetation index variables and other factors, so as to obtain reliable prediction for efficient production. Applying certain artificial intelligence algorithms to the precision agriculture can significantly improve the efficiency of traditional agriculture in crop planting and reduce the consumption of human and natural resources. Remote sensing can objectively, accurately and timely provide a large amount of information for ecological environment and crop growth in agriculture applications. By combining the image and spectral data obtained by remote sensing technology with machine learning, information about wheat growth, yield and insect pests can be learned in time.This thesis focuses on its applications in agriculture, particularly using effective prediction models such as the back propagation neural network and some optimisation algorithms for predicting wheat growth, yield and aphid. The work presented in this thesis address the issues of wheat growth prediction, yield assessment and aphid validation by model building and machine learning algorithm optimisation by means of remote sensing data. Specifically, the following objectives are defined: 1. Analyse multiple vegetation indexes based on the TM 1-4 band data of Landsat satellite and use regression algorithms to train the models and predict wheat growth; 2. Analyse and compare multiple vegetation indexes models by means of spectral data and use regression algorithms to predict wheat yield; 3. Combine spectral vegetation indexes and multiple regression algorithms to predict wheat aphid; 4. Use accurate evaluation criteria for validating the efficacy of the various algorithms.In this thesis, the remote sensing data from the satellite has been applied instead of the airborne-based remote sensing data. Based on the TM 1-4 band image data of Landsat satellite, multiple vegetation indexes were used as the input of regression algorithms. After that, four kinds of regression algorithms such as the multiple linear regression (MR) algorithm, back propagation network (BPNN) algorithm, genetic algorithm (GA) optimised BPNN algorithm and particle swarm optimisation (PSO) optimised BPNN algorithm were used to train the model and predict the LAI and SPAD. The prediction results of each algorithm were compared with the ground truth information collected by hand held instruments on the ground.The relationship between wheat yield and spectral data has been studied. Based on the BPNN algorithm, four kinds of models such as visible hyperspectral index (VHI) model, hyperspectral vegetation index (HVI) model, difference hyperspectral index (DHI) model and normalized hyperspectral index (NHI) model have been utilized to predict wheat yield. For the optimal NHI model, three regression algorithms such as back propagation network (BPNN) algorithm, genetic algorithm (GA) optimised BPNN algorithm and particle swarm optimisation (PSO) optimised BPNN algorithm, were compared to predict wheat yield, and RMSE and R-square of the three algorithms were compared and analysed.Finally, the relationship between wheat aphid and spectral data has been investigated. Nine vegetation indexes related to aphid have been estimated from spectral data as the input of regression algorithms. Five kinds of regression algorithms such as back propagation network (BPNN) algorithm, genetic algorithm (GA) optimised BPNN algorithm, particle swarm optimisation (PSO) optimised BPNN algorithm, ant colony (ACO) optimisation algorithm optimised BPNN algorithm and cuckoo search (CS) optimised BPNN algorithm have been implemented to predict wheat aphid, which was validated with the ground truth information measured by hand-held instruments on the ground. The prediction results of each algorithm have been analysed.The major original contributions of this thesis are as follows:1. A variety of optimisation algorithms are used to improve the regression analysis of the BPNN algorithm, so that the prediction results of each model for wheat growth, yield and aphid are more accurate.2. The spectral characteristics of winter wheat canopy have been analysed. The correlation between the absorption band and the associated physical and chemical properties of crops, specially the red edge slope, with the crop yield and wheat aphid damage is established.3. Adjusted MSE and un-centered R-square, as accurate evaluation criteria for practical applications, are used to compare the prediction results of the models under different dimensions of the observed data.4. Improve algorithm training by using the cross-validation method to obtain reliable and stable models for the prediction of wheat growth, yield, and aphid. Through repeated cross-validation, a better model can be obtained in the last.Key word:Precision agriculture; BP network, wheat growth assessment; wheat yield prediction, wheat aphid validation
Date of Award16 Mar 2020
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorJinchang Ren (Supervisor) & Hong Yue (Supervisor)

Cite this

'