The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-3/W12-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-3/W12-2020, 171–176, 2020
https://doi.org/10.5194/isprs-archives-XLII-3-W12-2020-171-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-3/W12-2020, 171–176, 2020
https://doi.org/10.5194/isprs-archives-XLII-3-W12-2020-171-2020

  06 Nov 2020

06 Nov 2020

UNSUPERVISED METHODOLOGY TO IN-SEASON MAPPING OF SUMMER CROPS IN URUGUAY WITH MODIS EVI’S TEMPORAL SERIES AND MACHINE LEARNING

A. Cal and G. Tiscornia A. Cal and G. Tiscornia
  • Unidad de Agroclima y Sistemas de Información (GRAS), Instituto Nacional de Investigación Agropecuaria (INIA), Uruguay

Keywords: crop mapping, elbow method, EVI, k-means, smoothing spline, time-series, unsupervised

Abstract. This paper presents a new methodology for mapping summer crops in Uruguay, during the season, based on time-series analysis of the EVI vegetation index derived from the MODIS sensor. Time-series were processed with the k-means unsupervised machine learning algorithm. For this algorithm, the ideal number of clusters was estimated using the elbow method. Once the clusters were obtained, for each one, the average phenological signature was adjusted using a nonlinear smoothing spline regression technique. Additionally, using the derivative analysis, the key points of the curve were estimated (minimum, maximum and inflection points). When analyzing the average signature of each cluster, those whose signature follows the seasonal pattern of an agricultural crop (similar to a Gaussian function) were selected to generate a binary map of crops/non-crops. The estimated crop area is 2,336,525 hectares, higher than the official statistics of 1,667,400 hectares for the 2014–15 season. This overestimation can be explained by the resolution of the MODIS pixel (250 meters), where each has a different degree of purity; and commission errors. The methodology was validated with 5,317 ground truth points, with a general accuracy of 95.8%, kappa index of 85.6, production and user accuracy of 85.1% and 91.3% for crops/non-crops.