The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-3/W6
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-3/W6, 585–592, 2019
https://doi.org/10.5194/isprs-archives-XLII-3-W6-585-2019
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-3/W6, 585–592, 2019
https://doi.org/10.5194/isprs-archives-XLII-3-W6-585-2019

  26 Jul 2019

26 Jul 2019

ASSESSING THE TRANSFERABILITY OF MACHINE LEARNING ALGORITHMS USING CLOUD COMPUTING AND EARTH OBSERVATION DATASETS FOR AGRICULTURAL LAND USE/COVER MAPPING

B. Praveen1, S. Mustak2, and Pritee Sharma1 B. Praveen et al.
  • 1Indian Institute of Technology, Indore, India
  • 2Ashoka Trust for Research in Ecology and Environment, Bangalore, India

Keywords: Cloud Computing, transferability, Kernel optimization, land use/land cover, earth observation

Abstract. Mapping of agricultural land use/cover was initiated since the past several decades for land use planning, change detection analysis, crop yield monitoring etc. using earth observation datasets and traditional parametric classifiers. Recently, machine learning, cloud computing, Google Earth Engine (GEE) and open source earth observation datasets widely used for fast, cost-efficient and precise agricultural land use/cover mapping and change detection analysis. Main objective of this study was to assess the transferability of the machine learning algorithms for land use/cover mapping using cloud computing and open source earth observation datasets. In this study, the Landsat TM (L5, L8) of 2018, 2009 and 1998 were selected and median reflectance of spectral bands in Kharif and Rabi season were used for the classification. In addition, three important machine learning algorithms such as Support Vector Machine with Radial Basis Function (SVM-RBF), Random forest (RF) and Classification and Regression Tree (CART) were selected to evaluate the performance in transferability for agricultural land use classification using GEE. Seven land use/cover classes such as built-up, cropland, fallow land, vegetation etc. were selected based on literature review and local land use classification scheme. In this classification, several strategies were employed such as feature extraction, feature selection, parameter tuning, sensitivity analysis on size of training samples, transferability analysis to assess the performance of the selected machine learning algorithms for land use/cover classification. The result shows that SVM-RBF outperforms the RF and CART for both spatial and temporal transferability analysis. This result is very helpful for agriculture and remote sensing scientist to suggest promising guideline to land use planner and policy-makers for efficient land use mapping, change detection analysis, land use planning and natural resource management.