Scholarlink

<p>Classification of satellite images using machine learning for crop map generation requires ground truth data representing diverse crop types. However, the use of ground truth collected across years is constrained due to inter-year variation in the cropping pattern in command areas with multi-cropping systems. Acquiring ground truth data is a cost and time-intensive process. This necessitates the need for a simplified method where labelled training data generated from collected ground truth can be extended subsequent years to generate crop maps throughout the growing season (referred to as in-season crop maps). This study compares five machine learning classifiers, Gradient Boosted Tree, Maximum Likelihood, Multi-Layer Perceptron Neural Network, Random Forest and Support Vector Machines for their performance in deriving in-season crop maps, using labelled training data generated from a non-concurrent ground truth. The primary input for crop classification is the surface reflectance at – 559.8&#xa0;nm, 664.6&#xa0;nm, 864.7&#xa0;nm, 1613.7&#xa0;nm, and 2202.4&#xa0;nm and vegetation indices. The classification performance was assessed using statistical metrics including Cohen’s Kappa, Overall Accuracy, and class-wise Precision, Recall, and F1-Score. To provide a holistic evaluation encompassing both overall and crop-specific classification efficacy, a Composite Performance Index (CPI) was derived by integrating these metrics. Random Forest classifier was found to perform better in generating in-season crop maps with better seasonal average value of Cohen’s Kappa (0.768), Overall Accuracy (84.08%) and CPI of 0.79.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Identification of Suitable Machine Learning Classifier for In-Season Crop Mapping in the Absence of Concurrent Ground Truth Data

  • 期刊论文

  • 发表日期：

    2025年12月4日

查看全文
  • Sanjay Mallya,
  • Annie Maria Issac,
  • S. Sithara,
  • Ronald Singh,
  • K. Abdul Hakeem,
  • K. Chandrasekar,
  • P. Venkat Raju

摘要

Classification of satellite images using machine learning for crop map generation requires ground truth data representing diverse crop types. However, the use of ground truth collected across years is constrained due to inter-year variation in the cropping pattern in command areas with multi-cropping systems. Acquiring ground truth data is a cost and time-intensive process. This necessitates the need for a simplified method where labelled training data generated from collected ground truth can be extended subsequent years to generate crop maps throughout the growing season (referred to as in-season crop maps). This study compares five machine learning classifiers, Gradient Boosted Tree, Maximum Likelihood, Multi-Layer Perceptron Neural Network, Random Forest and Support Vector Machines for their performance in deriving in-season crop maps, using labelled training data generated from a non-concurrent ground truth. The primary input for crop classification is the surface reflectance at – 559.8 nm, 664.6 nm, 864.7 nm, 1613.7 nm, and 2202.4 nm and vegetation indices. The classification performance was assessed using statistical metrics including Cohen’s Kappa, Overall Accuracy, and class-wise Precision, Recall, and F1-Score. To provide a holistic evaluation encompassing both overall and crop-specific classification efficacy, a Composite Performance Index (CPI) was derived by integrating these metrics. Random Forest classifier was found to perform better in generating in-season crop maps with better seasonal average value of Cohen’s Kappa (0.768), Overall Accuracy (84.08%) and CPI of 0.79.