Regional mapping of crops under agricultural nets using Sentinel-2

Date
2022-12
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: The use of agricultural nets is rapidly expanding worldwide as farmers are forced to adapt to the adverse effects of climate change on crops. These nets have diverse spectral properties due to their differing colours, thicknesses, porosities and the large variety of plastic compounds used to make the nets. As such, nets are difficult to map using existing remote sensing techniques. To address this problem, this study aimed to fill several research gaps pertaining to mapping agricultural nets, with a specific focus on the use of Sentinel-2 imagery. Sentinel-2 imagery is freely available and has a short (5-day) revisit time, making it a viable data source for monitoring large areas. However, there are limited research findings about whether Sentinel-2 imagery has the necessary spatial and spectral resolutions to effectively capture nets. This question was addressed through two experiments conducted in the Western Cape, South Africa. The first experiment aimed to record and interpret the spectral signatures of the most common types of nets used in the Western Cape and to investigate how these signatures are affected by the seasonal changes of their underlying crops. Spectral signatures of nets covering vineyards, citrus, and berry crops were collected for an entire growing season. The Jeffries-Matusita distance was used to quantify the spectral separability among the net classes and their surrounding land cover and how the signatures changed over time. The results showed that the spectral resolution of the Sentinel-2 imagery was adequate to identify distinguishable net signatures. Furthermore, it was found that the signatures of the underlying crops could also be identified. The net types were most separable during the summer months. The second experiment aimed to investigate whether Sentinel-2 imagery has the necessary spatial and spectral resolutions required to map nets. The experiment also tested different machine learning classifiers and classification features to determine which method and features were best suited for mapping nets. The classifications achieved high accuracies which were comparable to the accuracies achieved by studies that used very-high-resolution imagery (like WorldView-3 and QuickBird) for mapping agricultural nets. The accuracies achieved in the second experiment were higher than those found in studies using lower resolution Landsat imagery. The results showed that the unaltered Sentinel-2 bands contained the most important features for classification and that both the random forest and the neural network algorithms achieved high accuracies for mapping nets. Both experiments confirmed that Sentinel-2 imagery has the necessary spatial and spectral resolutions to effectively capture and map nets. This insight makes Sentinel-2 a practical and viable option for mapping strategies in agriculture. Furthermore, this study provided valuable information about the spectral characteristics of agricultural nets, and effective techniques for mapping the distribution of agricultural nets.
AFRIKAANSE OPSOMMING: Die gebruik van nette vir landbou neem wereldwyd toe, hoofsaaklik omdat hierdie nette die oeste teen die negatiewe gevolge van klimaatsverandering beskerm. Landbounette word van verskillende tipes plastiek gemaak en kan in vele kleure en digthede voorkom. Daarom het hierdie nette uiteenlopende spektrale eienskappe en word nette moeilik met die gebruik van satellietbeelde gekarteer. Om hierdie probleem op te los, ondersoek hierdie studie die gebruik van Sentinel-2 satellietbeelde om landbounette te klassifiseer. Sentinel-2 satellietbeelde is vrylik beskikbaar en word elke vyf dae herhaal. Hierdie twee eienskappe maak Sentinel-2 beelde geskik om groot landbougebiede met landbounette te karteer, maar ongelukkig was daar nog geen navorsing gedoen om Sentinel-2 beelde hiervoor te toets nie. Hierdie studie is gevolglik gerig om die navorsingsvraag te beantwoord. Om te bepaal of Sentinel-2 beelde vir die klassifikasie van landbounette gebruik kan word, het die huidige studie twee eksperimente uitgevoer. Die eerste eksperiment het die spektrale eienskappe van die landbounette ondersoek om te bepaal hoe hierdie nette in die Sentinel-2 beelde vertoon en om vas te stel of die spektrale eienskappe as gevolg van die plante onder die net oor tyd verander. ‘n Spektrale grafiek is vir drie tipes nette geskep, naamlik vir nette wat oor druiwe, sitrus en bessie gewasse gespan was. Die Jeffries-Matusita metode is gebruik om te bepaal of die unieke spektrale eienskappe van landbounette akkuraat genoeg in Sentinel-2 beelde uitgebeeld kan word sodat uitgekenning uitgevoer kan word. Daarna was die grafieke gebruik om vas te stel hoe die spektrale eienskappe van die landbounette verander soos wat die gewasse onder die nette gedurend die groeiseisoen verander. Die resultate toon dat die resolusie van die Sentinel-2 beelde goed genoeg is om die unieke spektrale eienskappe van die landbounette en die gewasse onder die nette vas te stel. Die resultate toon ook dat die landbounette die maklikste tydens somer geklassifiseer kan word. Die doel van die tweede eksperiment was om te sien of Sentinel-2 beelde vir die kartering van landbounette geskik is. Die eksperiment het ook ondersoek ingestel oor watter masjienleer metodes en veranderlikes die beste vir klassifikasie werk. Die neurale netwerk en ewekansige woud metodes het die beste resultate gelewer. Die akkuraatheid van die klassifikasies was hoer as kaarte wat deur ander navorsers met Landsat beelde gemaak is en net so hoog soos kaarte wat met baie hoe resolusie beelde (soos Wordview-3 en QuickBird) gemaak is. Die studie het bewys dat Sentinel-2 satellietbeelde gebruik kan word om landbounette te klassifiseer en om akkurate landboukaarte te maak. Verder voorsien die resultate unieke insig oor die spektrale eienskappe van landbounette, en watter masjienleer metodes en veranderlikes nuttig is om landbounette te klassifiseer.
Description
Thesis (MA)--Stellenbosch University, 2022.
Keywords
Agricultural netting, Jeffries-Matusita measure, Supervised learning (Machine learning), Neural networks (Computer science), Plastics in agriculture, Agricultural mapping, Sentinel-2, Landsat satellites, Hyperspectral imaging, Crop improvement, Digital image processing, Remote sensing, Climatic changes, UCTD
Citation