Selection of virtual remote sensing libraries and machine learning techniques for digitalimage processing applied to coffee crop

Abstract

Remote Sensing allows the possibility of monitoring and reasonably estimating the productivity, plant health and mineral nutrition of the coffee tree. By using sensors coupled to satellites it was possible to obtain information about the spectral signature of the coffee crop on a time scale relevant to the monitoring and detection of phenological changes. Surface reflectance can be obtained from a number of remote virtual sensing libraries, each of which can adopt a method of processing the data which may cause divergence in the information. The aim was to evaluate the source of orbital data acquisition and digital image processing for use in coffee growing. Data acquisition source was analyzed for radiometric and geometric differences between SATVeg, AppEEARS and Google Earth Engine platforms comparing a total of 900 sample points distributed in 3 scenes of MOD13Q1 for different dates. Regarding image processing, Machine Learning Random Forest, Naive Bayes and Rede Neural algorithms were evaluated to detect necrosis of coffee tree fruits by comparing the acuractia of the models using Friedman Nemenyi's method. To evaluate whether the Machine Learning algorithms are more effective than the agrometeorological spectral model, estimated productivity by both models using the time series of Landsat images. Based on the results the virtual platforms GEE and AppEEARS can be used with satisfactory accuracy regarding the radiometric values in the condition of being in the sinusoidal projection. Regarding the use of machine learning techniques to detect necrosis in coffee tree fruits, with the Naive Bayes method there were better results in the detection of fruit necrosis through Landsat images. On the yield estimation of coffee tree fruits, with the Random Forest method better estimates were observed in relation to the spectral agrometerological model, being a more indicated method when one intends to estimate the yield at pixel level of Landsat, in the area conditions and availability of images in which the experiment was performed.