Analýza satelitních snímků pro predikci výnosu plodin

Abstract

V tejto praci sa zameriavame na navrhnutie a implementaciu postupnosti krokov, ktora umozni predikciu urody plodin. V texte popisujeme a analyzujeme data pochadzajuce zo vzdialeneho prieskumu Zeme, ktore rozsirime o indexy vystihujuce vegetacne vlastnosti danej oblasti. Z tychto dat vyberieme podmnozinu urodnych poli. Tuto podmnozinu spojime s realnymi datami o urode a pouzijeme na natrenovanie a otestovanie regresnych modelov. V celom procese ma dolezitu ulohu platforma Google Earth Engine, ktora okrem pristupu k datam umoznuje aj nad nimi vykonavat rozne vypocty. V praci volime zakladne algoritmy strojoveho ucenia, ako algoritmus k-means, ci linearna regresia, so zamerom zistit, ci tieto zakladne metody su schopne dobrej predikcie. Vysledkom nasej prace je nastroj, ktory umoznuje predikciu urody. Model testujeme na predikcii urody zemiakov a obilnin. Vysledky testovania ukazuju, ze s predikciou obilnin si lepsie poradila kombinacia algoritmov Learning Vector Quantization a Support Vector Machine s absolutnou strednou chybou na urovni 0.2836 t ha[?]1. Pre urodu zemiakov nizsiu chybu, 5.3114 t ha[?]1, dosiahol algoritmus Learning Vector Quantization s linearnou regresiou.

In this work, we focus on suggesting and implementing a prediction pipeline which allows us to estimate a crop yield. We explore and analyse the Landsat remote sensing collection, extend it with indices that correlate with the vegetation level in order to extract cropland features. We associate these features with the actual crop yield values which are later used for training and testing a regression model. Google's Earth Engine platform plays an essential role in accessing the data and performing complex computations. As the extraction and prediction models, we choose basic machine learning approaches like k-means and Linear Regression with the intention of finding out if such models are capable of a good estimation. The result of our work is a tool which predicts crop yields. We test the models on cereals and potatoes datasets. The tests results show that Learning Vector Quantization - Support Vector Machine combination achieves the best results in the cereals dataset with Mean Absolute Error of 0.2836 t ha[?]1 and Learning Vector Quantization with Linear Regression in the potatoes dataset with Mean Absolute Error of 5.3114 t ha[?]1.