Fabian Löw

Defining the Spatial Resolution Requirements for Crop Identification Using Optical Remote Sensing

The past decades have seen an increasing demand for operational monitoring of crop conditions and food production at local to global scales. To properly use satellite Earth observation for such agricultural monitoring, high temporal revisit frequency over vast geographic areas is 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 question of determining the optimal pixel sizes for an application such as crop identification is therefore naturally inclined towards finding the coarsest acceptable pixel sizes, so as to potentially benefit from what instruments with coarser pixels can offer. To answer this question, this study builds upon and extends a conceptual framework to quantitatively define pixel size requirements for crop identification via 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. Results over contrasting landscapes in Central Asia demonstrate that the task of finding the optimum pixel size does not have a “one-size-fits-all” solution. The resulting values for pixel size and purity that are 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 crops are growing in. Results indicate that sensors like MODIS (250 m) could be suitable for identifying major crop classes in the study sites, whilst sensors like Landsat (30 m) should be considered for object-based classification. 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.

Per-field crop classification in irrigated agricultural regions in middle Asia using random forest and support vector machine ensemble

Accurate crop identification and crop area estimation are important for studies on irrigated agricultural systems, yield and water demand modeling, and agrarian policy development. In this study a novel combination of Random Forest (RF) and Support Vector Machine (SVM) classifiers is presented that (i) enhances crop classification accuracy and (ii) provides spatial information on map uncertainty. The methodology was implemented over four distinct irrigated sites in Middle Asia using RapidEye time series data. The RF feature importance statistics was used as feature-selection strategy for the SVM to assess possible negative effects on classification accuracy caused by an oversized feature space. The results of the individual RF and SVM classifications were combined with rules based on posterior classification probability and estimates of classification probability entropy. SVM classification performance was increased by feature selection through RF. Further experimental results indicate that the hybrid classifier improves overall classification accuracy in comparison to the single classifiers as well as user´s and producer´s accuracy.

Multi-Temporal Landsat Images and Ancillary Data for Land Use/Cover Change (LULCC) Detection in the Southwest of Burkina Faso, West Africa

Accurate quantification of land use/cover change (LULCC) is important for efficient environmental management, especially in regions that are extremely affected by climate variability and continuous population growth such as West Africa. In this context, accurate LULC classification and statistically sound change area estimates are essential for a better understanding of LULCC processes. This study aimed at comparing mono-temporal and multi-temporal LULC classifications as well as their combination with ancillary data and to determine LULCC across the heterogeneous landscape of southwest Burkina Faso using accurate classification results. Landsat data (1999, 2006 and 2011) and ancillary data served as input features for the random forest classifier algorithm. Five LULC classes were identified: woodland, mixed vegetation, bare surface, water and agricultural area. A reference database was established using different sources including high-resolution images, aerial photo and field data. LULCC and LULC classification accuracies, area and area uncertainty were computed based on the method of adjusted error matrices. The results revealed that multi-temporal classification significantly outperformed those solely based on mono-temporal data in the study area. However, combining mono-temporal imagery and ancillary data for LULC classification had the same accuracy level as multi-temporal classification which is an indication that this combination is an efficient alternative to multi-temporal classification in the study region, where cloud free images are rare. The LULCC map obtained had an overall accuracy of 92%. Natural vegetation loss was estimated to be 17.9% ± 2.5% between 1999 and 2011. The study area experienced an increase in agricultural area and bare surface at the expense of woodland and mixed vegetation, which attests to the ongoing deforestation. These results can serve as means of regional and global land cover products validation, as they provide a new validated data set with uncertainty estimates in heterogeneous ecosystems prone to classification errors.

The impact of training class proportions on binary cropland classification

ABSTRACT The ground truth data sets required to train supervised classifiers are usually collected as to maximize the number of samples under time, budget and accessibility constraints. Yet, the performance of machine learning classifiers is, among other factors, sensitive to the class proportions of the training set. In this letter, the joint effect of the number of calibration samples and the class proportions on the accuracy was systematically quantified using two state-of-the-art machine learning classifiers (random forests and support vector machines). The analysis was applied in the context of binary cropland classification and focused on two contrasted agricultural landscapes. Results showed that the classifiers were more sensitive to class proportions than to sample size, though sample size had to reach 2,000 pixels before its effect leveled off. Optimal accuracies were obtained when the training class proportions were close to those actually observed on the ground. Then, synthetic minority over-sampling technique (SMOTE) was implemented to artificially regenerate the native class proportions in the training set. This resampling method led to an increase of the accuracy of up to 30%. These results have direct implications for (i) informing data collection strategies and (ii) optimizing classification accuracy. Though derived for cropland mapping, the recommendations are generic to the problem of binary classification.

Potentials of RapidEye time series for improved classification of crop rotations in heterogeneous agricultural landscapes: Experiences from irrigation systems in Central Asia

In Central Asia, more than eight Million ha of agricultural land are under irrigation. But severe degradation problems and unreliable water distribution have caused declining yields during the past decades. Reliable and area-wide information about crops can be seen as important step to elaborate options for sustainable land and water management. Experiences from RapidEye classifications of crop in Central Asia are exemplarily shown during a classification of eight crop classes including three rotations with winter wheat, cotton, rice, and fallow land in the Khorezm region of Uzbekistan covering 230,000 ha of irrigated land. A random forest generated by using 1215 field samples was applied to multitemporal RapidEye data acquired during the vegetation period 2010. But RapidEye coverage varied and did not allow for generating temporally consistent mosaics covering the entire region. To classify all 55,188 agricultural parcels in the region three classification zones were classified separately. The zoning allowed for including at least three observation periods into classification. Overall accuracy exceeded 85 % for all classification zones. Highest accuracies of 87.4 % were achieved by including five spatiotemporal composites of RapidEye. Class-wise accuracy assessments showed the usefulness of selecting time steps which represent relevant phenological phases of the vegetation period. The presented approach can support regional crop inventory. Accurate classification results in early stages of the cropping season permit recalculation of crop water demands and reallocation of irrigation water. The high temporal and spatial resolution of RapidEye can be concluded highly beneficial for agricultural land use classifications in entire Central Asia.

Cropping Intensity in the Aral Sea Basin and Its Dependency from the Runoff Formation 2000-2012

This study is aimed at a better understanding of how upstream runoff formation affected the cropping intensity (CI: number of harvests) in the Aral Sea Basin (ASB) between 2000 and 2012. MODIS 250 m NDVI time series and knowledge-based pixel masking that included settlement layers and topography features enabled to map the irrigated cropland extent (iCE). Random forest models supported the classification of cropland vegetation phenology (CVP: winter/summer crops, double cropping, etc.). CI and the percentage of fallow cropland (PF) were derived from CVP. Spearman’s rho was selected for assessing the statistical relation of CI and PF to runoff formation in the Amu Darya and Syr Darya catchments per hydrological year. Validation in 12 reference sites using multi-annual Landsat-7 ETM+ images revealed an average overall accuracy of 0.85 for the iCE maps. MODIS maps overestimated that based on Landsat by an average factor of ~1.15 (MODIS iCE/Landsat iCE). Exceptional overestimations occurred in case of inaccurate settlement layers. The CVP and CI maps achieved overall accuracies of 0.91 and 0.96, respectively. The Amu Darya catchment disclosed significant positive (negative) relations between upstream runoff with CI (PF) and a high pressure on the river water resources in 2000–2012. Along the Syr Darya, reduced dependencies could be observed, which is potentially linked to the high number of water constructions in that catchment. Intensified double cropping after drought years occurred in Uzbekistan. However, a 10 km × 10 km grid of Spearman’s rho (CI and PF vs. upstream runoff) emphasized locations at different CI levels that are directly affected by runoff fluctuations in both river systems. The resulting maps may thus be supportive on the way to achieve long-term sustainability of crop production and to simultaneously protect the severely threatened environment in the ASB. The gained knowledge can be further used for investigating climatic impacts of irrigation in the region.

Assessing gaps in irrigated agricultural productivity through satellite earth observations—A case study of the Fergana Valley, Central Asia

Abstract Improving crop area and/or crop yields in agricultural regions is one of the foremost scientific challenges for the next decades. This is especially true in irrigated areas because sustainable intensification of irrigated crop production is virtually the sole means to enhance food supply and contribute to meeting food demands of a growing population. Yet, irrigated crop production worldwide is suffering from soil degradation and salinity, reduced soil fertility, and water scarcity rendering the performance of irrigation schemes often below potential. On the other hand, the scope for improving irrigated agricultural productivity remains obscure also due to the lack of spatial data on agricultural production (e.g. crop acreage and yield). To fill this gap, satellite earth observations and a replicable methodology were used to estimate crop yields at the field level for the period 2010/2014 in the Fergana Valley, Central Asia, to understand the response of agricultural productivity to factors related to the irrigation and drainage infrastructure and environment. The results showed that cropping pattern, i.e. the presence or absence of multi-annual crop rotations, and spatial diversity of crops had the most persistent effects on crop yields across observation years suggesting the need for introducing sustainable cropping systems. On the other hand, areas with a lower crop diversity or abundance of crop rotation tended to have lower crop yields, with differences of partly more than one t/ha yield. It is argued that factors related to the infrastructure, for example, the distance of farms to the next settlement or the density of roads, had a persistent effect on crop yield dynamics over time. The improvement potential of cotton and wheat yields were estimated at 5%, compared to crop yields of farms in the direct vicinity of settlements or roads. In this study it is highlighted how remotely sensed estimates of crop production in combination with geospatial technologies provide a unique perspective that, when combined with field surveys, can support planners to identify management priorities for improving regional production and/or reducing environmental impacts.

Mapping cropland abandonment in the Aral Sea Basin with MODIS time series

Cropland abandonment is globally widespread and has strong repercussions for regional food security and the environment. Statistics suggest that one of the hotspots of abandoned cropland is located in the drylands of the Aral Sea Basin (ASB), which covers parts of post-Soviet Central Asia, Afghanistan and Iran. To date, the exact spatial and temporal extents of abandoned cropland remain unclear, which hampers land-use planning. Abandoned land is a potentially valuable resource for alternative land uses. Here, we mapped the abandoned cropland in the drylands of the ASB with a time series of the Normalized Difference Vegetation Index (NDVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS) from 2003–2016. To overcome the restricted ability of a single classifier to accurately map land-use classes across large areas and agro-environmental gradients, “stratum-specific” classifiers were calibrated and classification results were fused based on a locally weighted decision fusion approach. Next, the agro-ecological suitability of abandoned cropland areas was evaluated. The stratum-specific classification approach yielded an overall accuracy of 0.879, which was significantly more accurate ( p < 0.05) than a “global” classification without stratification, which had an accuracy of 0.811. In 2016, the classification results showed that 13% (1.15 Mha) of the observed irrigated cropland in the ASB was idle (abandoned). Cropland abandonment occurred mostly in the Amudarya and Syrdarya downstream regions and was associated with degraded land and areas prone to water stress. Despite the almost twofold population growth and increasing food demand in the ASB area from 1990 to 2016, abandoned cropland was also located in areas with high suitability for farming. The map of abandoned cropland areas provides a novel basis for assessing the causes leading to abandoned cropland in the ASB. This contributes to assessing the suitability of abandoned cropland for food or bioenergy production, carbon storage, or assessing the environmental trade-offs and social constraints of recultivation.

Crop classification at subfield level using RapidEye time series and graph theory algorithms

Accurate information about land use patterns is crucial for a sustainable and economical use of water in agricultural systems. Water demand estimation, yield modeling and agrarian policy are only a few applications addressed by land use classifications based on remote sensing imagery. In Central Asia, where fields are traditionally large and state order crops dominate the area, small units of fields are often separated for the additional cultivation of income crops for the farmers. Traditional object based land use classifications on multi-temporal satellite imagery using field boundaries show low classification accuracies on these separated fields, expressed by a high uncertainty of the final class labels. Although segmentation of smaller subfields was shown to be suitable for improving the classification result, the extraction of subfields is still a time-consuming and error-prone process. In this study, energy based Graph-Cut segmentation technique is used to enhance the segmentation process and finally to improve the classification result. The interactive segmentation technique was successfully adopted from bio-medical image analysis to fit remote sensing imagery in the spatial and in the temporal domain. A set of rules was developed to perform the image segmentation procedure on pixels of single satellite datasets and on objects representing time series of a vegetation index. An ensemble classifier based on Random Forest and Support Vector Machines was used to receive information about classification uncertainty before and after applying the segmentation. It is demonstrated that subfield extraction based on Graph Cuts outperforms traditional image segmentation approaches in simplicity and reduces the risk of under- and over-segmentation significantly. Classification uncertainty decreased using the derived subfields as object boundaries instead of original field boundaries. The segmentation technique performs well on several multi-temporal satellite images without changing parameters and may be used to refine object based land use classifications to subfield level.

A rule-based approach for crop identification using multi-temporal and multi-sensor phenological metrics

ABSTRACT Accurate classification and mapping of crops is essential for supporting sustainable land management. Such maps can be created based on satellite remote sensing; however, the selection of input data and optimal classifier algorithm still needs to be addressed especially for areas where field data is scarce. We exploited the intra-annual variation of temporal signatures of remotely sensed observations and used prior knowledge of crop calendars for the development of a two-step processing chain for crop classification. First, Landsat-based time-series metrics capturing within-season phenological variation were preprocessed and analyzed using Google Earth Engine cloud computing platform. The developmental stage of each crop was modeled by fitting harmonic function. The model’s output was further used for the automatic generation of training samples. Second, several classification methods (support vector machines, random forest, decision fusion) were tested. As input data for crop classification, composites based on Sentinel-1 and Landsat images were used. Overall classification accuracies exceeded 80% when the seasonal composites were used. Winter cereals were the most accurately classified, while we observed misclassifications among summer crops. The proposed approach offers a potential to accurately map crops in the areas where in situ field data are scarce or unavailable.