Mapping cropping intensity of smallholder farms: A comparison of methods using multiple sensors

REMOTE SENSING OF ENVIRONMENT

卷:134

Article

Jain, Meha¹  Mondal, Pinki¹  DeFries, Ruth S.¹  Small, Christopher²  Galford, Gillian L.^1,3 

[1] Columbia Univ, Dept Ecol Evolut & Environm Biol, New York, NY 10027 USA

[2] Columbia Univ, Lamont Doherty Earth Observ, Palisades, NY 10964 USA

[3] Univ Vermont, Rubenstein Sch Environm & Nat Resources, Gund Inst Ecol Econ, Burlington, VT 05405 USA

关键词: Smallholder agriculture;    MODIS;    Landsat;    Cropping intensity;    Temporal mixture model;    Hierarchical training;    South Asia;   

DOI  :  10.1016/j.rse.2013.02.029

来源: Elsevier

【 摘 要 】

The food security of smallholder farmers is vulnerable to climate change and climate variability. Cropping intensity, the number of crops planted annually, can be used as a measure of food security for smallholder farmers given that it can greatly affect net production. Current techniques for quantifying cropping intensity may not accurately map smallholder farms where the size of one field is typically smaller than the spatial resolution of readily available satellite data. We evaluated four methods that use multi-scalar datasets and are commonly used in the literature to assess cropping intensity of smallholder farms: 1) the Landsat threshold method, which identifies if a Landsat pixel is cropped or uncropped during each growing season, 2) the MODIS peak method, which determines if there is a phenological peak in the MODIS Enhanced Vegetation Index time series during each growing season, 3) the MODIS temporal mixture analysis, which quantifies the sub-pixel heterogeneity of cropping intensity using phenological MODIS data, and 4) the MODIS hierarchical training method, which quantifies the sub-pixel heterogeneity of cropping intensity using hierarchical training techniques. Each method was assessed using four criteria: 1) data availability, 2) accuracy across different spatial scales (at aggregate scales 250 x 250 m, 1 x 1 km, 5 x 5 km, and 10 x 10 km), 3) ease of implementation, and 4) ability to use the method over large spatial and temporal scales. We applied our methods to two regions in India (Gujarat and southeastern Madhya Pradesh) that represented diversity in crop type, soils, climatology, irrigation access, cropping intensity, and field size. We found that the Landsat threshold method is the most accurate (R-2 >= 0.71 and RMSE <= 0.14), particularly at smaller scales of analysis. Yet given the limited availability of Landsat data, we find that the MODIS hierarchical training method meets multiple criteria for mapping cropping intensity over large spatial and temporal scales. Furthermore, the adjusted R-2 between predicted and validation data generally increased and the RMSE decreased with spatial aggregation >= 5 x 5 km (R-2 up to 0.97 and RMSE as low as 0.00). Our model accuracy varied based on the region and season of analysis and was lowest during the summer season in Gujarat when there was high sub-pixel heterogeneity due to sparsely cropped agricultural land-cover. While our results specifically apply to our study regions in India, they most likely also apply to smallholder agriculture in other locations across the globe where the same types of satellite data are readily available. (C) 2013 Elsevier Inc. All rights reserved.

【 授权许可】

Free