A quantitative performance comparison of paddy rice acreage estimation using stratified sampling strategies with different stratification indicators

Various indicators derived from thematic maps have been widely used to determine the strata needed to perform stratified sampling. However, these indicators typically do not quantify the spatial errors in the crop thematic maps that are needed to reduce the uncertainty. To address this lack of error information, this paper introduces a hybrid entropy indicator (HEI). Two conventional indicators, the acreage indicator (AI) and the fragmentation indicator (FI), were also evaluated to compare the results of the three indicators in a homogeneous agricultural area (Pinghu, PH) and a heterogeneous agricultural area (Zhuji, ZJ). The results show that HEI performs the best in heterogeneous areas with the lowest coefficient of variation (CV) (as low as 1.59%) and also has the highest estimation accuracy with the lowest standard deviation of estimation. For both areas, the performances of HEI and AI are very similar, and better than FI. These results highlight that the HEI should be considered as an effective indicator and used in place of AI and FI to help improve sampling efficiency of crop acreage estimation, while FI is not recommended. Furthermore, the positive performance achieved using HEI indicates the potential for incorporating thematic map uncertainty information to improve sampling efficiency.     

Using high resolution QuickBird imagery for crop identification and area estimation

Imagery from recently launched high spatial resolution satellite sensors offers new opportunities for crop assessment and monitoring. A 2.8-m multispectral QuickBird image covering an intensively cropped area in south Texas was evaluated for crop identification and area estimation. Three reduced-resolution images with pixel sizes of 11.2 m, 19.6 m, and 30.8 m were also generated from the original image to simulate coarser resolution imagery from other satellite systems. Supervised classification techniques were used to classify the original image and the three aggregated images into five crop classes (grain sorghum, cotton, citrus, sugarcane, and melons) and five non-crop cover types (mixed herbaceous species, mixed brush, water bodies, wet areas, and dry soil/roads). The five non-crop classes in the 10-category classification maps were then merged as one class. The classification maps were filtered to remove the small inclusions of other classes within the dominant class. For accuracy assessment of the classification maps, crop fields were ground verified and field boundaries were digitized from the original image to determine reference field areas for the five crops. Overall accuracy for the unfiltered 2.8-m, 11.2-m, 19.6-m, and 30.8-m classification maps were 71.4, 76.9, 77.1, and 78.0%, respectively, while overall accuracy for the respective filtered classification maps were 83.6, 82.3, 79.8, and 78.5%. Although increase in pixel size improved overall accuracy for the unfiltered classification maps, the filtered 2.8-m classification map provided the best overall accuracy. Percentage area estimates based on the filtered 2.8-m classification map (34.3, 16.4, 2.3, 2.2, 8.0, and 36.8% for grain sorghum, cotton, citrus, sugarcane, melons, and non-crop, respectively) agreed well with estimates from the digitized polygon map (35.0, 17.9, 2.4, 2.1, 8.0, and 34.6% for the respective categories). These results indicate that QuickBird imagery can be a useful data source for identifying crop types and estimating crop areas.     

Crop classification and acreage estimation in North Korea using phenology features

In North Korea, reliable and timely information on crop acreage and spatial distribution is hard to obtain. In this study, we developed a fast and robust method to estimate crop acreage in North Korea using time-series normalized difference vegetation index (NDVI) derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) data. We proposed a method to identify crop type based on NDVI phenology features using data collected in other areas with similar agri-environmental conditions to mitigate the shortage of ground truth data. Eventually the classification map (MODIScrop) was assessed using the Food and Agriculture Organization (FAO) statistical data and high-resolution crop classification maps derived from one Landsat scene (LScrop). The Pareto boundary method was used to assess the accuracy and crop distribution of the MODIScrop maps. Results showed that acreage derived from the MODIScrop maps was generally consistent with that reported in the FAO data (a relative error <4.1% for rice and <6.1% for maize, and <9.0% for soybean except for in 2004, 2008, and 2009) and the maps derived from the LScrop (a relative error about 5% in 2013, and 7% in 2008 and 2014). The classification accuracy reached 74.4%, 69.8%, and 73.1% of the areas covered by the Landsat images in 2008, 2013, and 2014, respectively. This indicates that features derived from NDVI profiles were able to characterize major crops, and the approaches developed in this study are feasible for crop mapping and acreage estimation in regions with limited ground truth data.     

Disaggregating and mapping crop statistics using hypertemporal remote sensing

Governments compile their agricultural statistics in tabular form by administrative area, which gives no clue to the exact locations where specific crops are actually grown. Such data are poorly suited for early warning and assessment of crop production. 10-Daily satellite image time series of Andalucia, Spain, acquired since 1998 by the SPOT Vegetation Instrument in combination with reported crop area statistics were used to produce the required crop maps. Firstly, the 10-daily (1998–2006) 1-km resolution SPOT-Vegetation NDVI-images were used to stratify the study area in 45 map units through an iterative unsupervised classification process. Each unit represents an NDVI-profile showing changes in vegetation greenness over time which is assumed to relate to the types of land cover and land use present. Secondly, the areas of NDVI-units and the reported cropped areas by municipality were used to disaggregate the crop statistics. Adjusted R-squares were 98.8% for rainfed wheat, 97.5% for rainfed sunflower, and 76.5% for barley. Relating statistical data on areas cropped by municipality with the NDVI-based unit map showed that the selected crops were significantly related to specific NDVI-based map units. Other NDVI-profiles did not relate to the studied crops and represented other types of land use or land cover. The results were validated by using primary field data. These data were collected by the Spanish government from 2001 to 2005 through grid sampling within agricultural areas; each grid (block) contains three 700 m × 700 m segments. The validation showed 68%, 31% and 23% variability explained (adjusted R-squares) between the three produced maps and the thousands of segment data. Mainly variability within the delineated NDVI-units caused relatively low values; the units are internally heterogeneous. Variability between units is properly captured. The maps must accordingly be considered “small scale maps”. These maps can be used to monitor crop performance of specific cropped areas because of using hypertemporal images. Early warning thus becomes more location and crop specific because of using hypertemporal remote sensing.     

Multi-Data Approach for remote sensing-based regional crop rotation mapping: A case study for the Rur catchment,Germany

Spatial land use information is one of the key input parameters for regional agro-ecosystem modeling. Furthermore, to assess the crop-specific management in a spatio-temporal context accurately, parcel-related crop rotation information is additionally needed. Such data is scarcely available for a regional scale, so that only modeled crop rotations can be incorporated instead. However, the spectrum of the occurring multiannual land use patterns on arable land remains unknown. Thus, this contribution focuses on the mapping of the actually practiced crop rotations in the Rur catchment, located in the western part of Germany. We addressed this by combining multitemporal multispectral remote sensing data, ancillary information and expert-knowledge on crop phenology in a GIS-based Multi-Data Approach (MDA). At first, a methodology for the enhanced differentiation of the major crop types on an annual basis was developed. Key aspects are (i) the usage of physical block data to separate arable land from other land use types, (ii) the classification of remote sensing scenes of specific time periods, which are most favorable for the differentiation of certain crop types, and (iii) the combination of the multitemporal classification results in a sequential analysis strategy. Annual crop maps of eight consecutive years (2008–2015) were combined to a crop sequence dataset to have a profound data basis for the mapping of crop rotations. In most years, the remote sensing data basis was highly fragmented. Nevertheless, our method enabled satisfying crop mapping results. As an example for the annual crop mapping workflow, the procedure and the result of 2015 are illustrated. For the generation of the crop sequence dataset, the eight annual crop maps were geometrically smoothened and integrated into a single vector data layer. The resulting dataset informs about the occurring crop sequence for individual areas on arable land, so that crop rotation schemes can be derived. The resulting dataset reveals that the spectrum of the practiced crop rotations is extremely heterogeneous and contains a large amount of crop sequences, which strongly diverge from model crop rotations. Consequently, the integration of remote sensing-based crop rotation data can considerably reduce uncertainties regarding the management in regional agro-ecosystem modeling. Finally, the developed methods and the results are discussed in detail.     

Remote sensing classification approach to large-scale crop cultivation identification: A case study of the Aral Sea Basin

Since the collapse of the Soviet Union, the crop cultivation structure in the Aral Sea Basin has changed dramatically, and these changes are worth studying. However, historical crop remote sensing mapping at the watershed scale remains challenging, especially crop misclassification at the cropland edge due to mixed pixels. Therefore, we proposed a field segmentation approach to constrain field edges based on time-series Sentinel-2 remote sensing images and the Google Earth Engine platform and then employed the random forest algorithm to perform crop classification based on time series Landsat/Sentinel-2 images and crop phenology information to produce historical crop maps in the Aral Sea Basin from the 1990s onward. The results showed that the intersection over union between the extracted field edges and in situ-measured field size data was 0.65. The overall accuracy of crop mapping was 95.2% in 2019. Then, we extended our method to historical mapping over the 1991–2015 period with accuracies ranging from 82.8% to 91.3%. Moreover, our method applied to historical mapping works well in terms of accuracy and policy matching. These findings indicate that our method can accurately distinguish cropland edges to reduce classification errors due to mixed pixels. This method is promising for solving the cropland edge problem for historical crop mapping in the Aral Sea Basin and can potentially provide a reference for historical crop classification in other watersheds of the world.     

土地覆盖变化信息自适应抽样策略及其精度评估

提出了一种面向土地覆盖变化信息局域精度评估的自适应型抽样策略。结合待研究地图的土地覆盖变化信息局部特征（如土地覆盖变化类别、斑块大小、异质性和优势度）,探讨与土地覆盖变化信息精度显著相关的协变量,以预测精度的标准误差作为判断标准,识别需要提高精度预测结果可靠性的区域,以自适应地和逐步定位的方式进行样本采集。基于武汉地区的精度评价结果,自适应的抽取增加100个训练样本使得预测精度的确定系数提高了50.66%,而简单随机抽取的增加样本使得预测精度的确定系数提高了17.22%。试验表明,自适应型抽样策略能显著提高土地覆盖变化信息局域精度预测的抽样效益,减少预测精度的不确定性。模型选择的结果表明,土地覆盖变化类别和优势度指数是最优的协变量组合。     

基于两个独立抽样框架的农作物种植面积遥感估算方法

通过分析遥感技术在中国农作物种植面积估算中所遇到的难点 ,针对运行化的农作物遥感估产系统对主要农作物种植面积估算的需求 ,提出在农作物种植结构区划的基础上 ,采用整群抽样和样条采样技术相结合的方法 ,进行农作物种植面积估算。整群抽样技术利用遥感影像估算农作物总种植成数 ,样条采样是一种适合中国农作物种植结构特征的采样技术 ,用于调查不同农作物类别在所有播种作物中的分类成数。在中国现有的耕地数据库基础上 ,根据两次抽样获得的成数 ,计算得到具体某一种农作物类别的种植面积。最后给出了 2 0 0 3年早稻种植面积估算的实例。     

A Spatial Conditioned Latin Hypercube Sampling Method for Mapping Using Ancillary Data

For obtaining maps of good precision by the spatial inference method, the distribution of sampling sites in geographical and feature space is very important. For a regional variable with trends, the predicting error comes from trend estimation, variogram estimation and spatial interpolation. Based on the cLHS (conditioned Latin hypercube Sampling) method, a sampling method called scLHS (spatial cLHS) considering all these three aspects with the help of ancillary data is proposed in this article. Its advantage lies in simultaneously improving trend estimation, variogram estimation and spatial interpolation. MODIS data and simulated data were used as sampling fields to draw sample sets using scLHS, cLHS, cLHS with x and y coordinates as covariates, simple random and spatial even sampling methods, and the distribution and prediction errors of sample sets from different methods were evaluated. The results showed that scLHS performed well in balancing spreading in geographic and feature space, and can generate points pairs with small distances, and the sample sets drawn by scLHS produced smaller mapping error, especially when there were trends in the target variable.     

Crop area mapping in West Africa using landscape stratification of MODIS time series and comparison with existing global land products

In Africa, food security early warning systems use satellite-derived data concerning crop conditions and agricultural production. Such systems can be improved if they are provided with a more reliable estimation of the cultivated area at national scale. This paper evaluates the potential of using time series from the MODerate resolution Imaging Spectroradiometer MOD13Q1 (16-day composite of normalized difference vegetation index at 250 m resolution) to extract cultivated areas in the fragmented rural landscapes of Mali. To this end, we first stratified Southern Mali into 13 rural landscapes based on the spatio-temporal variability of NDVI and textural indices, using an object-oriented classification scheme.The accuracy of the resulting map (MODIS_crop) and how it compares with existing coarse-resolution global land products (GLC2000 Africa, GLOBCOVER, MODIS V05 and ECOCLIMAP-II), was then assessed against six crop/non-crop maps derived from SPOT 2.5 m resolution images used as references. For crop areal coverage, the MODIS_crop cultivated map was successful in assessing the overall cultivated area at five out of the six validation sites (less than 6% of the absolute difference), while in terms of crop spatial distribution, the producer accuracy was between 33.1% and 80.8%. This accuracy was linearly correlated with the mean patch size index calculated on the SPOT crop maps (r² = 0.8). Using the Pareto boundary as an accuracy assessment method at the study sites, we showed that (i) 20-40% of the classification crop error was due to the spatial resolution of the MODIS sensor (250 m), and that (ii) compared to MODIS V05, which otherwise performed better than the other existing products, MODIS_crop generally minimized omission-commission errors. A spatial validation of the different products was carried out using SPOT image classifications as reference. In the corresponding error matrices, the fraction of correctly classified pixels for our product was 70%, compared to 58% for MODIS V05, while it ranged between 40% and 51% for the GLC2000, the ECOCLIMAP-II and the GLOBCOVER.     

基于对地抽样总量控制下的玉米种植面积提取

提出了一种基于统计抽样总量控制下的中高分辨率遥感影像玉米种植面积信息提取方法, 该方法首先利用分层抽样技术对调查目标总体(玉米)进行分层抽样;然后对抽样小区进行目视解译, 反推区域总量真值;最后在总量控制下进行区域目标作物的空间分布提取。以河北省三河市中部地区的部分影像为研究区, 以该区2006-08-21的10m分辨率的SPOT 5多光谱影像为基础数据进行了试验研究。结果表明该方法基于群样本检验的总体精度达到93.8%, Kappa系数达到0.88, 均高于最大似然监督分类结果的精度。另外, 所提出的方     

Efficient collection of training data for sub-pixel land cover classification using neural networks

Artificial neural networks (ANNs) are a popular class of techniques for performing soft classifications of satellite images. They have successfully been applied for estimating crop areas through sub-pixel classification of medium to low resolution images. Before a network can be used for classification and estimation, however, it has to be trained. The collection of the reference area fractions needed to train an ANN is often both time-consuming and expensive. This study focuses on strategies for decreasing the efforts needed to collect the necessary reference data, without compromising the accuracy of the resulting area estimates. Two aspects were studied: the spatial sampling scheme (i) and the possibility for reusing trained networks in multiple consecutive seasons (ii). Belgium was chosen as the study area because of the vast amount of reference data available. Time series of monthly NDVI composites for both SPOT-VGT and MODIS were used as the network inputs. The results showed that accurate regional crop area estimation (R² > 80%) is possible using only 1% of the entire area for network training, provided that the training samples used are representative for the land use variability present in the study area. Limiting the training samples to a specific subset of the population, either geographically or thematically, significantly decreased the accuracy of the estimates. The results also indicate that the use of ANNs trained with data from one season to estimate area fractions in another season is not to be recommended. The interannual variability observed in the endmembers’ spectral signatures underlines the importance of using up-to-date training samples. It can thus be concluded that the representativeness of the training samples, both regarding the spatial and the temporal aspects, is an important issue in crop area estimation using ANNs that should not easily be ignored.     

Farm-Level Yield mapping for precision crop management by linking remote sensing inputs and a crop simulation model

A study aimed at generating wheat yield maps of farmer’s fields by using remote sensing (RS) inputs was undertaken during the rabi season of 1998-99 in six villages of Alipur Block of Delhi State. RS derived leaf area index (LAI) were linked to wheat simulation model WTGROWS by adopting a strategy christened “Modified Corrective Approach”. This essentially uses an empirical relation of grain yield and LAI, which was derived from WTGROWS simulation model by running model for a combination of input resources, management practices and soil types occurring in the area. This biometric relationship was applied to all the wheat fields of the study area for which the LAI was derived from single acquisition of IRS LISS-III data (Jan 27, 99). The LAI-NDVI relation adopted was logarithmic in nature (R²=0.83) and was based on ground measurements of LAI in farmer’s fields in the same area. A comparison of predicted grain yield by the modified corrective approach and actual observed yield for the 22 farmer’s fields showed high correlation coefficient of 0.8 and a root mean square error (RMSE) of 597 kg ha^-1 which was 17% of the observed mean yield. Thus linking of RS information and crop simulation model provides an alternative for mapping and forecasting crop yield under highly variable cropping environment of Indian farms, which is a pre-requisite for implementing Precision Crop Management (PCM).     

MODIS植被指数的美国玉米单产遥感估测

针对中国开展的国外农作物产量遥感估测大多依靠中低分辨率耕地信息、省级(州级)或国家级作物产量统计数据的现状,本文以美国玉米为例,探讨利用多年中高分辨率作物分布信息、时序遥感植被指数和县级作物产量统计数据开展国外重点地区作物单产遥感估测技术研究,以期进一步提高中国对国外农作物产量监测精度和精细化水平。首先,利用美国农业部国家农业统计局(NASS/USDA)生产的作物分布数据(CDL)获得多个年份玉米空间分布图,并对相应年份250 m分辨率16天合成的MODIS-NDVI时序数据进行掩膜处理,统计获得每年各县域内玉米主要生育期NDVI均值;其次,以各州为估产区,以多年县级玉米统计单产和县域内玉米主要生育期NDVI均值为基础,建立各州玉米主要生育期NDVI与玉米单产间关系模型;然后,通过主要生育期玉米单产和玉米植被指数间拟合程度,筛选确定各州玉米最佳估产期和最佳估产模型。最终,利用最佳估产模型实现美国各州玉米单产估测和全国玉米单产推算。其中,建模数据覆盖时间为2007年—2010年,验证数据为2011年。结果表明,应用最佳估产模型的2011年美国各州玉米单产估测相对误差在-4.16%—4.92%,均方根误差在148.75—820.93 kg/ha,各州估测结果计算获得全国玉米单产的相对误差仅为2.12%,均方根误差为285.57 kg/ha。可见,本研究的作物单产遥感估测技术方法具有一定可行性,可准确估测全球重点地区作物单产信息。     

Use of landsat ETM+ SLC-off segment-based gap-filled imagery for crop type mapping

Failure of the Scan Line Corrector (SLC) on the Landsat ETM+ sensor has had a major impact on many applications that rely on continuous medium resolution imagery to meet their objectives. The United States Department of Agriculture (USDA) Cropland Data Layer (CDL) program uses Landsat imagery as the primary source of data to produce crop-specific maps for 20 states in the USA. A new method has been developed to fill the image gaps resulting from the SLC failure to support the needs of Landsat users who require coincident spectral data, such as for crop type mapping and monitoring. We tested the new gap-filled method for a CDL crop type mapping project in eastern Nebraska. Scan line gaps were simulated on two Landsat 5 images (spring and late summer 2003) and then gap-filled using landscape boundary models, or segment models, that were derived from 1992 and 2002 Landsat images (used in the gap-fill process). Various date combinations of original and gap-filled images were used to derive crop maps using a supervised classification process. Overall kappa values were slightly higher for crop maps derived from SLC-off gap-filled images compared to crop maps derived from the original imagery (0.3–1.3% higher). Although the age of the segment model used to derive the SLC-off gap-filled product did not negatively impact the overall agreement, differences in individual cover type agreement did increase (?0.8%–1.6% using the 2002 segment model to ?5.0–5.1% using the 1992 segment model). Classification agreement also decreased for most of the classes as the size of the segment used in the gap-fill process increased.     

Assessing the spatial distribution of crop areas using a cross-entropy method

While crop production statistics are reported on a geopolitical – often national – basis, we often need to know, for example, the status of production or productivity within specific sub-regions, watersheds, or agro-ecological zones. Such re-aggregations are typically made using expert judgments or simple area-weighting rules. We describe a new, entropy-based approach to the plausible estimates of the spatial distribution of crop areas. Using this approach tabular crop production statistics are blended judiciously with an array of other secondary data to assess the areas of specific crops within individual ‘pixels’—typically 25–100 km² in size. The information utilized includes crop production statistics, farming system characterization, satellite-based interpretation of land cover, biophysical crop suitability assessments, and population density. An application is presented in which Brazilian state level production statistics are used to generate pixel level crop area data for eight crops. To validate the spatial allocation we aggregated the pixel estimates to obtain synthetic estimates of municipality level areas in Brazil, and compared those estimates with actual municipality statistics. The approach produced extremely promising results. We then examined the robustness of these results compared to simplified approaches to spatializing crop production statistics and showed that, while computationally intensive, the cross-entropy method does provide more reliable spatial allocations.     

Using CORINE land cover and the point survey LUCAS for area estimation

CORINE land cover 2000 (CLC2000) is a European land cover map produced by photo-interpretation of Landsat ETM+ images. Its direct use for area estimation can be strongly biased and does not generally report single crops. CLC areas need to be calibrated to give acceptable statistical results.LUCAS (land use/cover area frame survey) is a point survey carried out in 2001 and 2003 in the European Union (EU15) on a systematic sample of clusters of points. LUCAS is especially useful for area estimation in geographic units that do not coincide with administrative regions, such as set of coastal areas defined with a 10 km buffer. Some variance estimation issues with systematic sampling of clusters are analysed.The contingency table obtained overlaying CLC and LUCAS gives the fine scale composition of CLC classes. Using CLC for post-stratification of LUCAS is equivalent to the direct calibration estimator when the sampling units are points. Stratification is easier to adapt to a scheme in which the sampling units are the clusters of points used in LUCAS 2001/2003.     

Phenological metrics-based crop classification using HJ-1 CCD images and Landsat 8 imagery

Crop type data are an important piece of information for many applications in agriculture. Extracting crop type using remote sensing is not easy because multiple crops are usually planted into small parcels with limited availability of satellite images due to weather conditions. In this research, we aim at producing crop maps for areas with abundant rainfall and small-sized parcels by making full use of Landsat 8 and HJ-1 charge-coupled device (CCD) data. We masked out non-vegetation areas by using Landsat 8 images and then extracted a crop map from a long-term time-series of HJ-1 CCD satellite images acquired at 30-m spatial resolution and two-day temporal resolution. To increase accuracy, four key phenological metrics of crops were extracted from time-series Normalized Difference Vegetation Index curves plotted from the HJ-1 CCD images. These phenological metrics were used to further identify each of the crop types with less, but easier to access, ancillary field survey data. We used crop area data from the Jingzhou statistical yearbook and 5.8-m spatial resolution ZY-3 satellite images to perform an accuracy assessment. The results show that our classification accuracy was 92% when compared with the highly accurate but limited ZY-3 images and matched up to 80% to the statistical crop areas.     

Satellite-based investigation of flood-affected rice cultivation areas in Chao Phraya River Delta,Thailand

The occurrence of catastrophic floods in Thailand in 2011 caused significant damage to rice agriculture. This study investigated flood-affected rice cultivation areas in the Chao Phraya River Delta (CRD) rice bowl, Thailand using time-series moderate resolution imaging spectroradiometer (MODIS) data. The data were processed for 2008 (normal flood year) and 2011, comprising four main steps: (1) data pre-processing to construct time-series MODIS vegetation indices (VIs), to filter noise from the time-series VIs by the empirical mode decomposition (EMD), and to mask out non-agricultural areas in respect to water-related cropping areas; (2) flood-affected area classification using the unsupervised linear mixture model (ULMM); (3) rice crop classification using the support vector machines (SVM); and (4) accuracy assessment of flood and rice crop mapping results. The comparisons between the flood mapping results and the ground reference data indicated an overall accuracy of 97.9% and Kappa coefficient of 0.62 achieved for 2008, and 95.7% and 0.77 for 2011, respectively. These results were reaffirmed by close agreement (R² > 0.8) between comparisons of the two datasets at the provincial level. The crop mapping results compared with the ground reference data revealed that the overall accuracies and Kappa coefficients obtained for 2008 were 88.5% and 0.82, and for 2011 were 84.1% and 0.76, respectively. A strong correlation was also found between MODIS-derived rice area and rice area statistics at the provincial level (R² > 0.7). Rice crop maps overlaid on the flood-affected area maps showed that approximately 16.8% of the rice cultivation area was affected by floods in 2011 compared to 4.9% in 2008. A majority of the flood-expanded area was observed for the double-cropped rice (10.5%), probably due to flood-induced effects to the autumn–summer and rainy season crops. Information achieved from this study could be useful for agricultural planners to mitigate possible impacts of floods on rice production.     

Impact of feature selection on the accuracy and spatial uncertainty of per-field crop classification using Support Vector Machines

Crop mapping is one major component of agricultural resource monitoring using remote sensing. Yield or water demand modeling requires that both, the total surface that is cultivated and the accurate distribution of crops, respectively is known. Map quality is crucial and influences the model outputs. Although the use of multi-spectral time series data in crop mapping has been acknowledged, the potentially high dimensionality of the input data remains an issue. In this study Support Vector Machines (SVM) are used for crop classification in irrigated landscapes at the object-level. Input to the classifications is 71 multi-seasonal spectral and geostatistical features computed from RapidEye time series. The random forest (RF) feature importance score was used to select a subset of features that achieved optimal accuracies. The relationship between the hard result accuracy and the soft output from the SVM is investigated by employing two measures of uncertainty, the maximum a posteriori probability and the alpha quadratic entropy. Specifically the effect of feature selection on map uncertainty is investigated by looking at the soft outputs of the SVM, in addition to classical accuracy metrics. Overall the SVMs applied to the reduced feature subspaces that were composed of the most informative multi-seasonal features led to a clear increase in classification accuracy up to 4.3%, and to a significant decline in thematic uncertainty. SVM was shown to be affected by feature space size and could benefit from RF-based feature selection. Uncertainty measures from SVM are an informative source of information on the spatial distribution of error in the crop maps.