Site-Specific Weed Management Using Remote Sensing

Abstract

The objective of this research was to develop a method to map weeds in sorghum as the firststep as a procedure to control them using site-specific weed management (SSWM). Site-specificweed management is a method to limit the application of herbicides only to areas with weeds.Accurate mapping of weeds is a pre-requisite for applying SSWM. Analysis of hyperspectralremote sensing imagery is recognized as a potentially cost effective technique for discriminatingbetween weeds and crop plants. This research involved: i) collecting hyperspectral reflectancespectra from weeds and sorghum plants, ii) Stepwise Linear Discriminant Analysis (SLDA) toidentify the most significant spectral bands, iii) Linear Discrimination Analysis (LDA) to test theaccuracy of the SLDA bands for classifying weeds and sorghum, and iv) analysis of customizedmultispectral imagery to produce maps which detected weeds in the sorghum crop.Hyperspectral signatures of weeds and sorghum were obtained using a FieldSpec®Handheld2TM spectroradiometer with a spectral range from 325 nm to 1075 nm. Spectra wererecorded for different weed species and sorghum plants for three years, 2012 to 2014. Data werecollected at four different stages of plant growth each year, from week one to week four afterplanting.The results show that it is feasible to discriminate spectral profiles of weeds from each otherweeds and from sorghum plants. Statistical Analysis Software (SAS) was used to identify the mostsignificant spectral bands (10 nm width) from the hyperspectral reflectance data using SLDA. Allweeds and sorghum were correctly classified in 2012 using LDA for week four reflectance data. In2013, the classification accuracy increased with stage of growth (weeks one to four) from 85% to90%. In 2014, the classification accuracy also increased with stage of growth (weeks two to four)from 90% to 100%. Combinations of spectral bands were analysed to reduce the number ofpotential bands identified from the SLDA results. Spectral bands centred on 930, 890, 710, 700, 560and 500 nm were common to the 20 most significant spectral bands identified by SLDA analysiseach year. Six spectral bands 850, 720, 710, 680, 560 and 440 nm were subsequently selected foruse in multispectral image collection. They were selected based on maximizing the differences andsimilarities between the 2013 weed and crop reflectance profiles. These bands were used for theband-pass filters in the Tetracam MCA 6 camera used for collecting high spatial resolutionterrestrial and aerial imagery.The imagery was analysed using Object-Based Image Analysis (OBIA) and VegetationIndex Analysis (VIA) to classify weeds and sorghum plants. OBIA identified weeds moresuccessfully than VIA. The accuracy of OBIA classification was tested using two methods.Confusion matrices were used to measure the Coefficient of Agreement (Khat), Overall Accuracy and Producer‘s and User Accuracies. Geometry matrices were used to measure under and oversegmentation. The Overall accuracy and Khat for all the weeds was more than 80% and 70%respectively for mosaic imagery. These results are considered high and moderate for effectivenessin discrimination respectively.The results of this research are limited by the weed species that grew in the sorghum crop atGatton, Queensland, the spectral resolution of the imagery and the image analysis methods. Toensure the wider applicability of the procedures, methods presented in this thesis need to be testedon other sorghum crops in different locations.