Accurate and timely spatial classification of crop types based on remote sensing data is important for both scientific and practical purposes. Spatially explicit crop-type information can be used to estimate crop areas for a variety of monitoring and decision-making applications such as crop insurance, land rental, supply-chain logistics, and financial market forecasting. However, there is no publically available spatially explicit in-season crop-type classification information for the U.S. Corn Belt (a landscape predominated by corn and soybean). Instead, researchers and decision-makers have to wait until four to six months after harvest to have such information from the previous year. The state-of-the-art research on crop-type classification has been shifted from relying on only spectral features of single static images to combining together spectral and time-series information. While Landsat data have a desirable spatial resolution for field-level crop-type classification, the ability to extract temporal phenology information based on Landsat data remains a challenge due to low temporal revisiting frequency and inevitable cloud contamination. To address this challenge and generate accurate, cost-effective, and in-season crop-type classification, this research uses the USDA's Common Land Units (CLUs) to aggregate spectral information for each field based on a time-series Landsat image data stack to largely overcome the cloud contamination issue while exploiting a machine learning model based on Deep Neural Network (DNN) and high-performance computing for intelligent and scalable computation of classification processes. Experiments were designed to evaluate what information is most useful for training the machine learning model for crop-type classification, and how various spatial and temporal factors affect the crop-type classification performance in order to derive timely crop type information. All experiments were conducted over Champaign County located in central Illinois, and a total of 1322 Landsat multi-temporal scenes including all the six optical spectral bands spanning from 2000 to 2015 were used. Computational experiments show the inclusion of temporal phenology information and evenly distributed spatial training samples in the study domain improves classification performance. The shortwave infrared bands show notably better performance than the widely used visible and near-infrared bands for classifying corn and soybean. In comparison with USDA's Crop Data Layer (CDL), this study found a relatively high Overall Accuracy (i.e. the number of the corrected classified fields divided by the number of the total fields) of 96% for classifying corn and soybean across all CLU fields in the Champaign County from 2000 to 2015. Furthermore, our approach achieved 95% Overall Accuracy by late July of the concurrent year for classifying corn and soybean. The findings suggest the methodology presented in this paper is promising for accurate, cost-effective, and in-season classification of field-level crop types, which may be scaled up to large geographic extents such as the U.S. Corn Belt.

Original languageEnglish (US)
Pages (from-to)35-47
Number of pages13
JournalRemote Sensing of Environment
StatePublished - Jun 1 2018


  • Crop type classification
  • Deep Neural Network (DNN)
  • Machine learning
  • Phenology
  • Remote sensing

ASJC Scopus subject areas

  • Soil Science
  • Geology
  • Computers in Earth Sciences


Dive into the research topics of 'A high-performance and in-season classification system of field-level crop types using time-series Landsat data and a machine learning approach'. Together they form a unique fingerprint.

Cite this