Search

Opazovanje zemlje iz vesolja je z napredki v razvoju satelitov mogoče v večji natančnosti in frekvenci zajema. S tem lahko ažurno in podrobno sledimo namembnosti in spremembi uporabe delov površja. Vendar pa je to mogoče le, če zajete podatke lahko dovolj hitro obdelamo. Pri tem je ključnega pomena razvoj metod strojnega učenja in računalniškega vida za samodejno razpoznavo. Na teh področjih v zadnjem času dominirajo metode globokega učenja. V magistrskem delu obravnavamo uporabo globokih nevronskih mrež za klasifikacijo tipa poljščin. Pri tem je ključnega pomena časovna informacija, zato najprej ovrednotimo dve arhitekturi nevronskih mrež, ki delujeta z uporabo meritev skozi čas. Kljub veliki količini prosto-dostopnih satelitskih podatkov je dostop do kvalitetnih učnih baz podatkov otežen, zaradi omejitev pri dostopnosti anotacij. Za potrebe eksperimentalne analize smo pripravili lastno bazo poljščin na področju Slovenije v letu 2017. Dostop do potrebnih podatkov smo imeli v okviru EU projekta Perceptive Sentinel. V podobnih študijah so bile baze omejene na manjše regije in po naših informacijah je to prva analiza, narejena na območju države. Pred učenjem metod smo ovrednotili pogostost časovnih meritev in tudi doprinos izbire spektralnih kanalov. Izbrano metodo smo nato nadgradili s konvolucijskimi plastmi, s katerimi smo zajeli tudi prostorski kontekst okoli obravnavane celice. Končna metoda še nekoliko izboljša povprečni rezultat na bazi, sicer zaradi fragmentiranosti zemljišč kontekst ni tako informativen, kot smo pričakovali, vendar pa so končne regije bolj homogene. With the advancement of remote sensing equipment, we have access to high frequency high resolution multi-spectral images of the world. This is a key component in up-to-date precision monitoring for changes in land use and land cover. But such feats can only be achieved if we can process the acquired data fast enough. This is why we require appropriate machine learning and computer vision algorithms. Nowadays these fields are dominated by deep learning and in our work we asses the use of deep neural networks for crop classification. An important component in this scenario is the use of temporal information. We experiment with two types of architectures that are capable of using temporal data. Despite the availability of satellite imagery, the constraints on high quality labeled data make training machine learning methods difficult. For the purpose of experiments in this work, we have prepared a dataset of crops in Slovenia during the year 2017. The access to the data was granted within the EU project Perceptive Sentinel. Similar studies were limited to smaller regions and to the best of our knowledge this will be the first one done on a whole country. We first evaluate the importance of the time series frequency and compare the importance of spectral information. The chosen method is then further improved with the use of spatial information, which enables us to include the context around the observed cell. The method achieves better average performance on the selected domain, but due to the high fragmentation of the fields in the dataset the improvement is not as large as one would have expected, however, the resulting regions are more homogeneous.