Retro-Classification of Agricultural Land Use in Irrigated lowlands of the Amu Darya River Catchment in Uzbekistan
Dubovyk, Olena1; Conrad, Christopher2; Khamzina, Asia3; Thonfeld, Frank4; Menz, Gunter1
1University of Bonn, GERMANY; 2University of Würzburg, GERMANY; 3Center for Development Research, GERMANY; 4Center for Remote Sensing of Land Surfaces, GERMANY

Site-specific crop identification is important for agro-ecological studies, regional climate modeling, and agricultural policy development. Moreover, availability of past crop maps is necessary for the understanding of land management practices and their effectiveness, as well as monitoring of environmental impacts of land uses, such as land degradation. In general in Central Asia and in Uzbekistan in particular, major land use transformations have happened in the past decades after the collapse of the Soviet Union in 1991. Lack of accurate information on current and past land use forestalls the assessment of occurred changes and their consequences, thus complicating the development of knowledge-driven agrarian policies. At the same time, lack of the sampling dataset for the past years often restricts mapping of past land use. In heterogeneous irrigated agro-ecosystems of Uzbekistan, conventional pixel-based analysis of satellite data could lead to inaccurate cropland mapping due to mixed pixels, pixel heterogeneity, and spectral similarity of crops. In contrast, object-based image analysis (OBIA) can handle these issues, at the same time providing parcel-specific land use information. We combined OBIA and Random Forest decision tree algorithms to develop a methodology for a retro-assessment of several crops, using a limited sampling dataset. In our approach, we explored several vegetation estimates, i.e., vegetation indices, vegetation and soil fraction images, Tasseled Cap indices, and textural features, derived from visible and near-infrared bands of multiple Landsat TM images that were collected during three distinct crop growing seasons (early summer, midsummer, and early fall) in years 1987, 1998, 2009, and 2010. The classification model, which was built for the main crops (cotton, winter wheat, rice, and tree plantations) and their intra-annual rotations for 2010 by using independent training and validation datasets, resulted in overall accuracy of 86%. The derived model was subsequently applied to classify years 1987, 1998, and 2009 after performing pair-wise image normalization using the Multivariate Alteration Detection algorithm based on year 2010. To validate the proposed approach, we transferred the classification model from 2010 back to the images of 2009 (''retro-classification'') and calculated the error matrix, using independent validation dataset collected in 2009. The validation yielded an overall accuracy of 80%, while commission and emission errors ranged between 8% (rice) and 42% (trees). Next, the validated classification model was applied to map cropping patterns in 1987 and 1998. The classification results revealed the main crop changes between analyzed years, e.g.., the area under cotton decreased from 132,260 ha in 1987 to 93,751 ha in 2010, while area under winter wheat and its rotations increased from 29,538 ha in 1987 to 79,471 ha in 2010. The increased area under winter wheat confirms the state policy for increased yield quotas for this crop for national self-sufficiency in wheat, introduced in 1990s. The proposed method enabled detecting and differentiating the dominant crops and their rotations in the irrigated lowlands of the Amu Darya River catchment with sufficient overall accuracies. However, it was not possible to map accurately the less prevalent crops, such as maize and sunflower, probably due to the limited number of available samples in 2010. Spectral variables (based on vegetation indices) contributed mostly to the classification models, although textural variables were crucial to discriminate among most crops. The proposed method was built using transferable rules, and it was successful for object-based retro-classification of agricultural land use in irrigated lowlands of Central Asia.