Semi-Supervised Technique to Retrieve Irrigated Crops from Landsat ETM+ Imagery for Small Fields and Mixed Cropping Systems of South Asia.
要旨Precise land cover maps and crop statistics are essential for monitoring and performance evaluation of any irrigation system. Remote sensing techniques are efficient means of obtaining land cover information with high accuracy in near real time. Many applications developed and referenced in the literature to classify large tracts of homogenous cropped area, although they may be less effective in highly mixed cropping systems in South Asia. This paper presents a semi-supervised (hybrid) technique to map cropped area in the Punjab Rice-Wheat and Sugarcane-Wheat cropping systems in Rechna Doab, Pakistan. Two Landsat 7 ETM+ images (Path 149 Row 038) for September 2001 and March 2002 were selected to represent Kharif (2001) and Rabi (2001-02) seasons. A ground truth survey was conducted in different cropping areas that covered 0.14% of the total extent, which was used to prepare a geo-referenced GIS coverage for image classification. Based on Principle Component Analy sis (PCA), Red, Near Infrared (NIR) and Mid Infrared (MIR) bands were selected, with and Normalized Difference Vegetation Index (NDVI), to segregate cropping area based on threshold values. GIS coverages were overlaid on a stacked image and 50% of the collected information was used to extract training signatures for all crops. The training samples were evaluated based on their location and overlap on two scatter plots between NIR: Red and NIR: MIR which yielded distinct crop classes. A contingency matrix was prepared using the other half of the field data to evaluate the classification accuracy. The results in Rechna Doab show an overall accuracy of 87% in Kharif (Summer-Monsoon) compared to 83% in Rabi (Winter-Dry). Inter-cropping of wheat with orchards and signature mixing with Fodder results comparatively low overall accuracy in Rabi despite the lower diversity of crops within the season. The present study shows how NDVI, when combined with individual bands improves the s ensitivity and accuracy of classification.