AGRICULTURAL LAND IMAGE CLASSIFICATION USING KNN AND COMPARE WITH RECURRENT NEURAL NETWORK

Abstract

Research into agriculture has been exhibiting indications of fast expansion over the last
several years. The most recent entrant is making farming easier by using a variety of
computational technology. We have used LAND satellite photos, which include
coverage of FOREST, AGRICULTURE, URBAN, and RANGING LAND, to execute
this project. While several classifiers have been developed for use with Sentinel-2
Multispectral Imager (MSI) and similar remote sensing pictures, very few research have
examined their capabilities using various training sample sizes. In this work, we used
Sentinel-2 picture data to classify land use and cover using RF, kNN, and SVM
classifiers. We compared their results. A total of fourteen alternative training sample
sizes, ranging from fifty pixels per class to more than twelve hundred, were used to
categorise a thirty by thirty km2 region in the Red River Delta of Vietnam that had six
distinct land use/cover classes. These sizes included balanced and unbalanced options.
A high overall accuracy (OA) between 90% and 95% was shown by all categorisation
findings. Using the training sample sizes as little a factor as possible, SVM generated
the greatest OA across all three classifiers and fourteen sub-datasets. RNN and kNN
trailed closely behind. With a sufficiently enough training sample size (i.e., more than
750 pixels/class or around 0.25 percent of the overall study area), all three classifiers
showed a comparable and high OA. With both balanced and unbalanced datasets, the
high accuracy was reached.