Soil classification with hyper-temporal satellite images

Abstract

Soil classification is a classic topic in the field of modern soil research. With the development of remote sensing technology, accurate soil classification can be predicted using different temporal satellite data. Therefore, it is crucial to compare the temporal response of satellite images. In this paper, the study area chosen was in Mingshui County of northeastern China, which is known as a black soil area, and a total of 34 years of Landsat satellite images were obtained, ranging from 1984 to 2018. We extracted six reflectance bands and three tasseled cap transformation indexes from each image as characteristics, and constructed a random forest (RF) model for classification. The results indicated that we could accurately predict the spatial distribution of the soil classes using hyper-temporal satellite images with an overall accuracy (OA) of 80.56% and a Kappa coefficient of 0.704. Compared with the mono-, bi-, and multi-temporal data, the overall accuracy using hyper-temporal data was increased by 17.78%, 12.23%, and 8.89%, and the Kappa coefficient using hyper-temporal data was increased by 0.265, 0.185, and 0.136, respectively. Our study provides a new perspective for updating the legacy soil data.