@phdthesis{Knauer2018,
  author    = {Knauer, Kim},
  title     = {Vegetation Dynamics in West Africa - Spatio-temporal Data Fusion for the Monitoring of Agricultural Expansion},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-164776},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2018},
  abstract  = {West Africa is one of the fastest growing regions in the world with annual population growth rates of more than three percent for several countries. Since the 1950s, West Africa experienced a fivefold increase of inhabitants, from 71 to 353 million people in 2015 and it is expected that the region's population will continue to grow to almost 800 million people by the year 2050. This strong trend has and will have serious consequences for food security since agricultural productivity is still on a comparatively low level in most countries of West Africa. In order to compensate for this low productivity, an expansion of agricultural areas is rapidly progressing. The mapping and monitoring of agricultural areas in West Africa is a difficult task even on the basis of remote sensing. The small scale extensive farming practices with a low level of agricultural inputs and mechanization make the delineation of cultivated land from other land cover and land use (LULC) types highly challenging. In addition, the frequent cloud coverage in the region considerably decreases the availability of earth observation datasets. For the accurate mapping of agricultural area in West Africa, high temporal as well as spatial resolution is necessary to delineate the small-sized fields and to obtain data from periods where different LULC types are distinguishable. However, such consistent time series are currently not available for West Africa. Thus, a spatio-temporal data fusion framework was developed in this thesis for the generation of high spatial and temporal resolution time series. Data fusion algorithms such as the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) enjoyed increasing popularity during recent years but they have hardly been used for the application on larger scales. In order to make it applicable for this purpose and to increase the input data availability, especially in cloud-prone areas such as West Africa, the ESTARFM framework was developed in this thesis introducing several enhancements. An automatic filling of cloud gaps was included in the framework in order to use even partly cloud-covered Landsat images for the fusion without producing gaps on the output images. In addition, the ESTARFM algorithm was improved to automatically account for regional differences in the heterogeneity of the study region. Further improvements comprise the automation of the time series generation as well as the significant acceleration of the processing speed through parallelization. The performance of the developed ESTARFM framework was tested by fusing an 8-day NDVI time series from Landsat and MODIS data for a focus area of 98,000 km² in the border region between Burkina Faso and Ghana. The results of this test show the capability of the ESTARFM framework to accurately produce high temporal resolution time series while maintaining the spatial detail, even in such a heterogeneous and cloud-prone region. The successfully tested framework was subsequently applied to generate consistent time series as the basis for the mapping of agricultural area in Burkina Faso for the years 2001, 2007, and 2014. In a first step, high temporal (8-day) and high spatial (30 m) resolution NDVI time series for the entire country and the three years were derived with the ESTARFM framework. More than 500 Landsat scenes and 3000 MODIS scenes were automatically processed for this purpose. From the fused ESTARFM NDVI time series, phenological metrics were extracted and together with the single time steps of NDVI served as input for the delineation of rainfed agricultural areas, irrigated agricultural areas and plantations. The classification was conducted with the random forest algorithm at a 30 m spatial resolution for entire Burkina Faso and the three years 2001, 2007, and 2014. For the training and validation of the classifier, a randomly sampled reference dataset was generated from Google Earth images based on expert knowledge of the region. The overall classification accuracies of 92\% (2001), 91\% (2007), and 91\% (2014) indicate the well-functioning of the developed methodology. The resulting maps show an expansion of agricultural area of 91\% from about 61,000 km² in 2001 to 116,900 km² in 2014. While rainfed agricultural areas account for the major part of this increase, irrigated areas and plantations also spread considerably. Especially the expansion of irrigation systems and plantation area can be explained by the promotion through various national and international development projects. The increase of agricultural areas goes in line with the rapid population growth in most of Burkina Faso's provinces which still had available land resources for an expansion of agricultural area. An analysis of the development of agricultural areas in the vicinity of protected areas highlighted the increased human pressure on these reserves. The protection of the remnant habitats for flora and fauna while at the same time improving food security for a rapidly growing population, are the major challenges for the region in the future. The developed ESTARFM framework showed great potential beyond its utilization for the mapping of agricultural area. Other large-scale research that requires a sufficiently high temporal and spatial resolution such as the monitoring of land degradation or the investigation of land surface phenology could greatly benefit from the application of this framework.},
  subject      = {Fernerkundung},
  language  = {en}
}
@phdthesis{Loew2013,
  author    = {L{\"o}w, Fabian},
  title     = {Agricultural crop mapping from multi-scale remote sensing data - Concepts and applications in heterogeneous Middle Asian agricultural landscapes},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-102093},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2013},
  abstract  = {Agriculture is mankind's primary source of food production and plays the key role for cereal supply to humanity. One of the future challenges will be to feed a constantly growing population, which is expected to reach more than nine billion by 2050. The potential to expand cropland is limited, and enhancing agricultural production efficiency is one important means to meet the future food demand. Hence, there is an increasing demand for dependable, accurate and comprehensive agricultural intelligence on crop production. The value of satellite earth observation (EO) data for agricultural monitoring is well recognized. One fundamental requirement for agricultural monitoring is routinely updated information on crop acreage and the spatial distribution of crops. With the technical advancement of satellite sensor systems, imagery with higher temporal and finer spatial resolution became available. The classification of such multi-temporal data sets is an effective and accurate means to produce crop maps, but methods must be developed that can handle such large and complex data sets. Furthermore, to properly use satellite EO for agricultural production monitoring a high temporal revisit frequency over vast geographic areas is often necessary. However, this often limits the spatial resolution that can be used. The challenge of discriminating pixels that correspond to a particular crop type, a prerequisite for crop specific agricultural monitoring, remains daunting when the signal encoded in pixels stems from several land uses (mixed pixels), e.g. over heterogeneous landscapes where individual fields are often smaller than individual pixels. The main purposes of the presented study were (i) to assess the influence of input dimensionality and feature selection on classification accuracy and uncertainty in object-based crop classification, (ii) to evaluate if combining classifier algorithms can improve the quality of crop maps (e.g. classification accuracy), (iii) to assess the spatial resolution requirements for crop identification via image classification. Reporting on the map quality is traditionally done with measures that stem from the confusion matrix based on the hard classification result. Yet, these measures do not consider the spatial variation of errors in maps. Measures of classification uncertainty can be used for this purpose, but they have attained only little attention in remote sensing studies. Classifier algorithms like the support vector machine (SVM) can estimate class memberships (the so called soft output) for each classified pixel or object. Based on these estimations, measures of classification uncertainty can be calculated, but it has not been analysed in detail, yet, if these are reliable in predicting the spatial distribution of errors in maps. In this study, SVM was applied for the classification of agricultural crops in irrigated landscapes in Middle Asia at the object-level. Five different categories of features were calculated from RapidEye time series data as classification input. The reliability of classification uncertainty measures like entropy, derived from the soft output of SVM, with regard to predicting the spatial distribution of error was evaluated. Further, the impact of the type and dimensionality of the input data on classification uncertainty was analysed. The results revealed that SMVs applied to the five feature categories separately performed different in classifying different types of crops. Incorporating all five categories of features by concatenating them into one stacked vector did not lead to an increase in accuracy, and partly reduced the model performance most obviously because of the Hughes phenomena. Yet, applying the random forest (RF) algorithm to select a subset of features led to an increase of classification accuracy of the SVM. The feature group with red edge-based indices was the most important for general crop classification, and the red edge NDVI had an outstanding importance for classifying crops. Two measures of uncertainty were calculated based on the soft output from SVM: maximum a-posteriori probability and alpha quadratic entropy. Irrespective of the measure used, the results indicate a decline in classification uncertainty when a dimensionality reduction was performed. The two uncertainty measures were found to be reliable indicators to predict errors in maps. Correctly classified test cases were associated with low uncertainty, whilst incorrectly test cases tended to be associated with higher uncertainty. The issue of combining the results of different classifier algorithms in order to increase classification accuracy was addressed. First, the SVM was compared with two other non-parametric classifier algorithms: multilayer perceptron neural network (MLP) and RF. Despite their comparatively high classification performance, each of the tested classifier algorithms tended to make errors in different parts of the input space, e.g. performed different in classifying crops. Hence, a combination of the complementary outputs was envisaged. To this end, a classifier combination scheme was proposed, which is based on existing algebraic operators. It combines the outputs of different classifier algorithms at the per-case (e.g. pixel or object) basis. The per-case class membership estimations of each classifier algorithm were compared, and the reliability of each classifier algorithm with respect to classifying a specific crop class was assessed based on the confusion matrix. In doing so, less reliable classifier algorithms were excluded at the per-class basis before the final combination. Emphasis was put on evaluating the selected classification algorithms under limiting conditions by applying them to small input datasets and to reduced training sample sets, respectively. Further, the applicability to datasets from another year was demonstrated to assess temporal transferability. Although the single classifier algorithms performed well in all test sites, the classifier combination scheme provided consistently higher classification accuracies over all test sites and in different years, respectively. This makes this approach distinct from the single classifier algorithms, which performed different and showed a higher variability in class-wise accuracies. Further, the proposed classifier combination scheme performed better when using small training set sizes or when applied to small input datasets, respectively. A framework was proposed to quantitatively define pixel size requirements for crop identification via image classification. That framework is based on simulating how agricultural landscapes, and more specifically the fields covered by one crop of interest, are seen by instruments with increasingly coarser resolving power. The concept of crop specific pixel purity, defined as the degree of homogeneity of the signal encoded in a pixel with respect to the target crop type, is used to analyse how mixed the pixels can be (as they become coarser) without undermining their capacity to describe the desired surface properties (e.g. to distinguish crop classes via supervised or unsupervised image classification). This tool can be modulated using different parameterizations to explore trade-offs between pixel size and pixel purity when addressing the question of crop identification. Inputs to the experiments were eight multi-temporal images from the RapidEye sensor. Simulated pixel sizes ranged from 13 m to 747.5 m, in increments of 6.5 m. Constraining parameters for crop identification were defined by setting thresholds for classification accuracy and uncertainty. Results over irrigated agricultural landscapes in Middle Asia demonstrate that the task of finding the optimum pixel size did not have a "one-size-fits-all" solution. The resulting values for pixel size and purity that were suitable for crop identification proved to be specific to a given landscape, and for each crop they differed across different landscapes. Over the same time series, different crops were not identifiable simultaneously in the season and these requirements further changed over the years, reflecting the different agro-ecological conditions the investigated crops were growing in. Results further indicate that map quality (e.g. classification accuracy) was not homogeneously distributed in a landscape, but that it depended on the spatial structures and the pixel size, respectively. The proposed framework is generic and can be applied to any agricultural landscape, thereby potentially serving to guide recommendations for designing dedicated EO missions that can satisfy the requirements in terms of pixel size to identify and discriminate crop types. Regarding the operationalization of EO-based techniques for agricultural monitoring and its application to a broader range of agricultural landscapes, it can be noted that, despite the high performance of existing methods (e.g. classifier algorithms), transferability and stability of such methods remain one important research issue. This means that methods developed and tested in one place might not necessarily be portable to another place or over several years, respectively. Specifically in Middle Asia, which was selected as study region in this thesis, classifier combination makes sense due to its easy implementation and because it enhanced classification accuracy for classes with insufficient training samples. This observation makes it interesting for operational contexts and when field reference data availability is limited. Similar to the transferability of methods, the application of only one certain kind of EO data (e.g. with one specific pixel size) over different landscapes needs to be revisited and the synergistic use of multi-scale data, e.g. combining remote sensing imagery of both fine and coarse spatial resolution, should be fostered. The necessity to predict and control the effects of spatial and temporal scale on crop classification is recognized here as a major goal to achieve in EO-based agricultural monitoring.},
  subject      = {Fernerkundung},
  language  = {en}
}