Landsat-based operational wheat area estimation model for Punjab, Pakistan

Abstract

Wheat in Punjab province of Pakistan is grown during the Rabi (winter) season within a heterogeneous smallholder agricultural system subject to a range of pressures including water scarcity, climate change and variability, and management practices. Punjab is the breadbasket of Pakistan, representing over 70% of national wheat production. Timely estimation of cultivated wheat area can serve to inform decision-making in managing harvests with regard to markets and food security. The current wheat area and yield reporting system, operated by the Punjab Crop Reporting Service (CRS) delivers crop forecasts several months after harvest. The delayed production data cannot contribute to in-season decision support systems. There is a need for an alternative cost-effective, efficient and timely approach on producing wheat area estimates, in ensuring food security for the millions of people in Pakistan. Landsat data, medium spatial and temporal resolutions, offer a data source for characterizing wheat in smallholder agriculture landscapes. This dissertation presents methods for operational mapping of wheat cultivate area using within growing season Landsat time-series data. In addition to maps of wheat cover in Punjab, probability-based samples of in-situ reference data were allocated using the map as a stratifier. A two-stage probability based cluster field sample was used to estimate area and assess map accuracies. The before-harvest wheat area estimates from field-based sampling and Landsat map were found to be comparable to official post-harvest data produced by the CRS Punjab. This research concluded that Landsat medium resolution data has sufficient spatial and temporal coverage for successful wall-to-wall mapping of wheat in Punjab’s smallholder agricultural system. Freely available coarse and medium spatial resolution satellite data such as MODIS and Landsat perform well in characterizing industrial cropping systems; commercial high spatial resolution satellite data are often advocated as an alternative for characterizing fine-scale land tenure agricultural systems such as that found in Punjab. Commercial 5 m spatial resolution RapidEye data from the peak of the winter wheat growing season were used as sub-pixel training data in mapping wheat with the growing season free 30 m Landsat time series data from the 2014-15 growing season. The use of RapidEye to calibrate mapping algorithms did not produce significantly higher overall accuracies ( ± standard error) compared to traditional whole pixel training of Landsat-based 30 m data. Continuous wheat mapping yielded an overall accuracy of 88% (SE = ±4%) in comparison to 87% (SE = ±4%) for categorical wheat mapping, leading to the finding that sub-pixel training data are not required for winter wheat mapping in Punjab. Given sufficient expertise in supervised classification model calibration, freely available Landsat data are sufficient for crop mapping in the fine-scale land tenure system of Punjab. For winter wheat mapping in Punjab and other similar landscapes, training data for supervised classification may be collected directly from Landsat images with probability based stratified random sampling as reference data without the need for high-resolution reference imagery. The research concluded by exploring the use of automated models in wheat area mapping and area estimation using growing season Landsat time-series data. The automated classification tree model resulted in wheat / not wheat maps with comparable accuracies compared to results achieved with traditional manual training. In estimating area, automated wheat maps from previous growing seasons can serve as a stratifier in the allocation of current season in-situ reference data, and current growing season maps can serve as an auxiliary variable in model-assisted area estimation procedures. The research demonstrated operational implementation of robust automated mapping in generating timely, accurate, and precise wheat area estimates. Such information is a critical input to policy decisions, and can help to ensure appropriate post-harvest grain management to address situations arising from surpluses or shortages in crop production.