Classification and Segmentation of Agricultural Land Using Linear Discriminant Analysis for Soil Sensors Installation

Abstract

An urgent problem of the soil state monitoring is solved using specific sensors. This rises another specific problem of search appropriate places on the terrain, which can be fulfilled on aerial or satellite images. The search process can be automated using computer vision techniques such as classification, segmentation, etc. This requires appropriate features, that represent significant properties of the terrain. Multisperspectral indices such as NDVI, NDBI, NDMI, SAVI, etc. represent various properties of the landcover on the satellite images. However, the established values and intervals of these indices may not always represent the proper properties of the terrain, because of various distortion factors. Thus it is necessary to perform additional statistical analysis in order to approve the difference in the considering indices for various classes. This work is devoted to statistical analysis of NDVI and NDBI and defining the statistically justified intervals of these indices for the following terrain classes on the satellite images, that may be suitable (class Soil) and unsuitable (classes Swamp, Water, Urban Objects, Forest) for soil sensors installation. The data is collected from various regions in Russia from Sentinel-2 images. Mann-Whitney U test has shown statistically significant difference in all the considering classes (p<0.001). The intervals of the considering indices have been obtain using Linear Discriminant Anaysis (LDA). The results of segmentation using established criteria for the Soil class is 0.63 according to Intersection Over Union (IoU) metric. Classification of Soil class has shown 0.76, 0.78 and 0.77 according to recall, presision and f1-score metrics correspondingly.