Enhanced land use/cover classification of heterogeneous tropical landscapes using support vector machines and textural homogeneity

Land use/cover classification is a key research field in remote sensing and land change science as thematic maps derived from remotely sensed data have become the basis for analyzing many socio-ecological issues. However, land use/cover classification remains a difficult task and it is especially challenging in heterogeneous tropical landscapes where nonetheless such maps are of great importance. The present study aims at establishing an efficient classification approach to accurately map all broad land use/cover classes in a large, heterogeneous tropical area, as a basis for further studies (e.g., land use/cover change, deforestation and forest degradation). Specifically, we first compare the performance of parametric (maximum likelihood), non-parametric (k-nearest neighbor and four different support vector machines – SVM), and hybrid (unsupervised–supervised) classifiers, using hard and soft (fuzzy) accuracy assessments. We then assess, using the maximum likelihood algorithm, what textural indices from the gray-level co-occurrence matrix lead to greater classification improvements at the spatial resolution of Landsat imagery (30 m), and rank them accordingly. Finally, we use the textural index that provides the most accurate classification results to evaluate whether its usefulness varies significantly with the classifier used. We classified imagery corresponding to dry and wet seasons and found that SVM classifiers outperformed all the rest. We also found that the use of some textural indices, but particularly homogeneity and entropy, can significantly improve classifications. We focused on the use of the homogeneity index, which has so far been neglected in land use/cover classification efforts, and found that this index along with reflectance bands significantly increased the overall accuracy of all the classifiers, but particularly of SVM. We observed that improvements in producer's and user's accuracies through the inclusion of homogeneity were different depending on land use/cover classes. Early-growth/degraded forests, pastures, grasslands and savanna were the classes most improved, especially with the SVM radial basis function and SVM sigmoid classifiers, though with both classifiers all land use/cover classes were mapped with producer's and user's accuracies of ∼90%. Our classification approach seems very well suited to accurately map land use/cover of heterogeneous landscapes, thus having great potential to contribute to climate change mitigation schemes, conservation initiatives, and the design of management plans and rural development policies.     

Mapping the distribution of the main host for plague in a complex landscape in Kazakhstan: An object-based approach using SPOT-5 XS,Landsat 7 ETM+, SRTM and multiple Random Forests

Plague is a zoonotic infectious disease present in great gerbil populations in Kazakhstan. Infectious disease dynamics are influenced by the spatial distribution of the carriers (hosts) of the disease. The great gerbil, the main host in our study area, lives in burrows, which can be recognized on high resolution satellite imagery. In this study, using earth observation data at various spatial scales, we map the spatial distribution of burrows in a semi-desert landscape.The study area consists of various landscape types. To evaluate whether identification of burrows by classification is possible in these landscape types, the study area was subdivided into eight landscape units, on the basis of Landsat 7 ETM+ derived Tasselled Cap Greenness and Brightness, and SRTM derived standard deviation in elevation.In the field, 904 burrows were mapped. Using two segmented 2.5 m resolution SPOT-5 XS satellite scenes, reference object sets were created. Random Forests were built for both SPOT scenes and used to classify the images. Additionally, a stratified classification was carried out, by building separate Random Forests per landscape unit.Burrows were successfully classified in all landscape units. In the ‘steppe on floodplain’ areas, classification worked best: producer's and user's accuracy in those areas reached 88% and 100%, respectively. In the ‘floodplain’ areas with a more heterogeneous vegetation cover, classification worked least well; there, accuracies were 86 and 58% respectively. Stratified classification improved the results in all landscape units where comparison was possible (four), increasing kappa coefficients by 13, 10, 9 and 1%, respectively.In this study, an innovative stratification method using high- and medium resolution imagery was applied in order to map host distribution on a large spatial scale. The burrow maps we developed will help to detect changes in the distribution of great gerbil populations and, moreover, serve as a unique empirical data set which can be used as input for epidemiological plague models. This is an important step in understanding the dynamics of plague.     

Comparing supervised and unsupervised multiresolution segmentation approaches for extracting buildings from very high resolution imagery

Although multiresolution segmentation (MRS) is a powerful technique for dealing with very high resolution imagery, some of the image objects that it generates do not match the geometries of the target objects, which reduces the classification accuracy. MRS can, however, be guided to produce results that approach the desired object geometry using either supervised or unsupervised approaches. Although some studies have suggested that a supervised approach is preferable, there has been no comparative evaluation of these two approaches. Therefore, in this study, we have compared supervised and unsupervised approaches to MRS. One supervised and two unsupervised segmentation methods were tested on three areas using QuickBird and WorldView-2 satellite imagery. The results were assessed using both segmentation evaluation methods and an accuracy assessment of the resulting building classifications. Thus, differences in the geometries of the image objects and in the potential to achieve satisfactory thematic accuracies were evaluated. The two approaches yielded remarkably similar classification results, with overall accuracies ranging from 82% to 86%. The performance of one of the unsupervised methods was unexpectedly similar to that of the supervised method; they identified almost identical scale parameters as being optimal for segmenting buildings, resulting in very similar geometries for the resulting image objects. The second unsupervised method produced very different image objects from the supervised method, but their classification accuracies were still very similar. The latter result was unexpected because, contrary to previously published findings, it suggests a high degree of independence between the segmentation results and classification accuracy. The results of this study have two important implications. The first is that object-based image analysis can be automated without sacrificing classification accuracy, and the second is that the previously accepted idea that classification is dependent on segmentation is challenged by our unexpected results, casting doubt on the value of pursuing ‘optimal segmentation’. Our results rather suggest that as long as under-segmentation remains at acceptable levels, imperfections in segmentation can be ruled out, so that a high level of classification accuracy can still be achieved.     

Comparing object-based and pixel-based classifications for mapping savannas

The development of robust object-based classification methods suitable for medium to high resolution satellite imagery provides a valid alternative to ‘traditional’ pixel-based methods. This paper compares the results of an object-based classification to a supervised per-pixel classification for mapping land cover in the tropical north of the Northern Territory of Australia. The object-based approach involved segmentation of image data into objects at multiple scale levels. Objects were assigned classes using training objects and the Nearest Neighbour supervised and fuzzy classification algorithm. The supervised pixel-based classification involved the selection of training areas and a classification using the maximum likelihood classifier algorithm. Site-specific accuracy assessment using confusion matrices of both classifications were undertaken based on 256 reference sites. A comparison of the results shows a statistically significant higher overall accuracy of the object-based classification over the pixel-based classification. The incorporation of a digital elevation model (DEM) layer and associated class rules into the object-based classification produced slightly higher accuracies overall and for certain classes; however this was not statistically significant over the object-based using spectral information solely. The results indicate object-based analysis has good potential for extracting land cover information from satellite imagery captured over spatially heterogeneous land covers of tropical Australia.     

Employing airborne multispectral digital imagery to map Brazilian pepper infestation in South Texas

A study was conducted in south Texas to determine the feasibility of using airborne multispectral digital imagery for differentiating the invasive plant Brazilian pepper (Schinus terebinthifolius) from other cover types. Imagery obtained in the visible, near-infrared, and mid-infrared regions of the light spectrum and a supervised classification approach were employed to develop thematic maps of two areas infested with Brazilian pepper. Map accuracies ranged from 84.2 to 100% for the Brazilian pepper class. Findings support using airborne multispectral digital imagery as a tool for separating Brazilian pepper from associated land cover types and further encourage exploration of airborne multispectral digital imagery and image processing techniques for developing maps of Brazilian pepper infestation in Texas and abroad.     

The Utility of Very High Spatial Resolution Images to Identify Urban Objects

Abstract

Landsat MSS, TM and SPOT XS imageries were used in conjunction with unsupervised, supervised and hybrid classilication techniques to classify land cover types in semi‐arid savannas of Mathison Pastoral Station in the Katherine region of northern Australia. Accuracy assessment was based on field data from 246 ground survey sites over a 745‐km² study area. Of 14 land cover classes identified by traditional mapping means, all combinations of imageries and classification techniques differentiated at least seven land cover types. The overall accuracy for these classifications ranged between 43% and 67%. SPOT XS image delivered the best accuracy followed by TM and MSS; unsupervised classification performed better than supervised and hybrid methods. User's and producer's accuracy of individual land units ranged from 0% to 100%. Riparian woodlands, woodland on limestone slopes, shrubland on clay plains, woodland on limestone plains and shadows were the best‐mapped classes. The land units that were associated with undulating hills were not mapped accurately. However, incorporation of a digital elevation model (DEM) in a GIS improved the overall accuracy. The user's and producer's accuracy of dominant land cover types were also enhanced. The classification results and the efficacy of the techniques at Mathison were similar to those found for a nearby semi‐arid area (Kidman Springs) about 200 km from Mathison. However, the overall accuracy was lower at Mathison than at Kidman Springs. Spectral classification masks were developed from the SPOT XS and TM imageries at Kidman Springs, and were applied to classify SPOT XS and TM imageries at Mathison. Initial results showed that the classification mask could be successfully extrapolated to map dominant land cover types but only with moderate accuracy (50%).     

通过训练样本采样处理改善小宗作物遥感识别精度

训练样本质量是决定农作物遥感识别精度的关键因素,虽然高空间分辨率卫星的发展有效地解决了农作物遥感识别过程中的混合像元问题,但是当区域内不同作物种植面积差异较大时,训练集中不同类别样本数量往往相差较大,这样的不均衡数据集影响分类器的训练,导致少数类别的识别精度不理想。为研究作物遥感识别过程中的不均衡样本问题,本文基于GF-2号卫星数据,首先挖掘了地物的光谱信息、纹理信息,用特征递归消除RFE (Recursive Feature Elimination)方法进行特征优选,然后从数据处理的角度采用了5种采样算法对不均衡训练集进行处理,最后使用采样后的均衡数据集训练分类器,对比数据采样前后决策树与Adaboost(Adaptive Boosting)两种分类器的识别结果,发现:(1)经过采样处理后两种分类算法明显提升了小宗作物的分类精度;(2)经过ADASYS (Adaptive synthetic sampling)采样处理后,分类器性能提升最多,决策树的Kappa系数提高了14.32%,Adaboost的Kappa系数提高了10.23%,达到最高值0.9336;(3)过采样的处理效果优于欠采样,过采样对分类器的性能提升更多。综上所述,选择合适的采样方法和分类方法是提高不均衡数据集遥感分类精度的有效途径。     

A multi-resolution satellite imagery approach for large area mapping of ericaceous shrubs in Northern Quebec, Canada

Invasive ericaceous shrubs (e.g. Kalmia angustifolia, Rhododendron groenlandicum, Vaccinium spp.) may reduce the regeneration and early growth of black spruce (Picea mariana) seedlings, the most economically important boreal tree species in Quebec. Our study focused, therefore, on developing a method for mapping ericaceous shrubs from satellite images. The method integrates very high resolution satellite imagery (IKONOS) to guide classifiers applied to medium resolution satellite imagery (Landsat-TM). An object-oriented image classification approach was applied using Definiens eCognition software. An independent ground survey revealed 80% accuracy at the very high spatial resolution. We found that the partial use (70%) of classified polygons derived from the IKONOS images were an effective way to guide classification algorithms applied to the Landsat-TM imagery. The results of this latter classification (78.4% overall accuracy) were assessed by the remaining portion (30%) of unused very high resolution classified polygons. We further validated our method (65.5% overall accuracy) by assessing the correspondence of an ericaceous cover classification scheme done with a Landsat-TM image and results of our ground survey using an independent set of 275 sample plots. Discrimination of ericaceous shrub cover from other land cover types was achieved with precision at both spatial resolutions with producer accuracies of 87.7% and 79.4% from IKONOS and Landsat, respectively. The method is weaker for areas with sparse cover of ericaceous shrubs or dense tree cover. Our method is adapted, therefore, for mapping the spatial distribution of ericaceous shrubs and is compatible with existing forest stand maps.     

The georeferencing of RASAT satellite imagery and some practical approaches to increase the georeferencing accuracy

RASAT Earth Observation Satellite is the second remote sensing satellite of The Scientific and Technological Research Council of Turkey (TUBITAK) Space Technologies Search Institute (TUBITAK Space). Generally, the first step to utilize the satellite imagery in GIS applications is the accurate geometric correction of the imagery. But, the geometric correction process of RASAT images is somewhat difficult due to insufficient orbit data and lack of satellite imagery processing software with RASAT model. Although the geolocation of RASAT images is investigated in some recent studies, subpixel accuracies cannot be achieved. In this study, different geometric correction methods and combination of them are tested with a more interactive workflow that uses the results of other approaches. Results show that these approaches can be used for the geometric correction of RASAT like pushbroom satellite images with insufficient orbit data and better geolocation accuracies can be achieved by different geometric correction approaches.     

European forest cover mapping with high resolution satellite data: The Carpathians case study

The aim of the study was to elaborate a methodology for forest mapping based on high resolution satellite data, relevant for reporting on forest cover and spatial pattern changes in Europe. The Carpathians were selected as a case study area and mapped using 24 Landsat scenes, processed independently with a supervised approach combining image segmentation, knowledge-based rules to extract a training set and the maximum likelihood decision rule. Validation was done with available very high resolution imagery. Overall accuracies per scene ranged from 93 to 96%. The labelling disagreement in overlapping areas of adjacent scenes was 6.8% on average.     

Remote classification of Cerrado (Savanna) and agricultural land covers in northeastern Brazil

Remote classification of land-use/land-cover (LULC) types in Brazil's Cerrado ecoregion is necessary because knowledge of Cerrado LULC is incomplete, sources of inaccuracy are unknown, and high-resolution data are required for the validation of moderate-resolution LULC maps. The aim of this research is to discriminate between Cerrado and agriculture using high-resolution Landsat 7 ETM⁺ imagery for the western region of Bahia state in northeastern Brazil. The Maximum Likelihood Classification (MLC) and Spectral Angle Mapper (SAM) algorithms were applied to a ～3000 km² subset, yielding comparable classification accuracies. The panchromatic band was reserved for validation. User's and producer's accuracies were highest for non-irrigated agriculture (～94%) but lower for Cerrado Lato Sensu (89%). Classification errors likely resulted from spatial and spectral characteristics of particular classes (e.g. riparian forest and burned) and overestimation of other classes (e.g. Eucalyptus and water). Manual misinterpretation of validation data may have also led to lower reported classification accuracies.     

Spatial application of Random Forest models for fine-scale coastal vegetation classification using object based analysis of aerial orthophoto and DEM data

High spatial resolution mapping of natural resources is much needed for monitoring and management of species, habitats and landscapes. Generally, detailed surveillance has been conducted as fieldwork, numerical analysis of satellite images or manual interpretation of aerial images, but methods of object-based image analysis (OBIA) and machine learning have recently produced promising examples of automated classifications of aerial imagery. The spatial application potential of such models is however still questionable since the transferability has rarely been evaluated.We investigated the potential of mosaic aerial orthophoto red, green and blue (RGB)/near infrared (NIR) imagery and digital elevation model (DEM) data for mapping very fine-scale vegetation structure in semi-natural terrestrial coastal areas in Denmark. The Random Forest (RF) algorithm, with a wide range of object-derived image and DEM variables, was applied for classification of vegetation structure types using two hierarchical levels of complexity. Models were constructed and validated by cross-validation using three scenarios: (1) training and validation data without spatial separation, (2) training and validation data spatially separated within sites, and (3) training and validation data spatially separated between different sites.Without spatial separation of training and validation data, high classification accuracies of coastal structures of 92.1% and 91.8% were achieved on coarse and fine thematic levels, respectively. When models were applied to spatially separated observations within sites classification accuracies dropped to 85.8% accuracy at the coarse thematic level, and 81.9% at the fine thematic level. When the models were applied to observations from other sites than those trained upon the ability to discriminate vegetation structures was low, with 69.0% and 54.2% accuracy at the coarse and fine thematic levels, respectively.Evaluating classification models with different degrees of spatial correlation between training and validation data was shown to give highly different prediction accuracies, thereby highlighting model transferability and application potential. Aerial image and DEM-based RF models had low transferability to new areas due to lack of representation of aerial image, landscape and vegetation variation in training data. They do, however, show promise at local scale for supporting conservation and management with vegetation mappings of high spatial and thematic detail based on low-cost image data.     

Mapping trees in high resolution imagery across large areas using locally variable thresholds guided by medium resolution tree maps

Large area tree maps, important for environmental monitoring and natural resource management, are often based on medium resolution satellite imagery. These data have difficulty in detecting trees in fragmented woodlands, and have significant omission errors in modified agricultural areas. High resolution imagery can better detect these trees, however, as most high resolution imagery is not normalised it is difficult to automate a tree classification method over large areas. The method developed here used an existing medium resolution map derived from either Landsat or SPOT5 satellite imagery to guide the classification of the high resolution imagery. It selected a spatially-variable threshold on the green band, calculated based on the spatially-variable percentage of trees in the existing map of tree cover. The green band proved more consistent at classifying trees across different images than several common band combinations. The method was tested on 0.5 m resolution imagery from airborne digital sensor (ADS) imagery across New South Wales (NSW), Australia using both Landsat and SPOT5 derived tree maps to guide the threshold selection. Accuracy was assessed across 6 large image mosaics revealing a more accurate result when the more accurate tree map from SPOT5 imagery was used. The resulting maps achieved an overall accuracy with 95% confidence intervals of 93% (90–95%), while the overall accuracy of the previous SPOT5 tree map was 87% (86–89%). The method reduced omission errors by mapping more scattered trees, although it did increase commission errors caused by dark pixels from water, building shadows, topographic shadows, and some soils and crops. The method allows trees to be automatically mapped at 5 m resolution from high resolution imagery, provided a medium resolution tree map already exists.     

基于BP神经网络的高分辨率遥感影像分类

为了能有效地从高分辨率遥感影像中提取地物信息,本文通过影像的光谱和纹理特征,利用BP神经网络算法进行影像分类研究。首先提取分类所需的光谱和纹理特征源,然后根据影像和地物特征,建立BP神经网络,用于样本训练和分类处理,实现地物分类。为验证该方法的可靠性,以2006年11月获取的成都平原某区域的Quickbird影像为实验数据,进行高分辨率遥感影像的地物分类实验。实验结果表明,结合影像光谱和纹理特征的BP神经网络分类算法,不仅可以有效保证BP神经网络分类训练的稳定性和收敛速度,还能达到较高的分类精度。     

直线特征约束的高分辨率卫星影像区域网平差方法

以"像方观测直线与像方预测直线必须重合"作为几何约束条件,以有理函数模型(RFM)作为高分辨率卫星影像的几何处理模型,提出了一种直线特征约束的高分辨率卫星影像区域网平差方法。本文方法仅需像方直线与物方直线相对应,无须像方直线上的像点与物方直线上的地面点一一对应。通过对圣迭戈试验区的两景IKONOS影像、斯波坎试验区的两景QuickBird影像和普罗旺斯试验区的两景SPOT-5影像进行试验,结果表明:本文方法可以充分利用直线特征作为控制条件,有效补偿RPC参数中的系统误差,获得的IKONOS、QuickBird和SPOT-5影像区域网平差的平面与高程精度均优于1个像素。     

基于稀疏判别分析的高光谱影像特征提取

针对当前特征提取方法不能充分挖掘高光谱影像稀疏特性的问题,提出一种基于稀疏判别分析的高光谱影像特征提取方法。首先,在线性判别分析的系数向量中引入稀疏正则项来捕获具有更强判别能力的特征,将高光谱影像映射至低维稀疏的子空间;然后,利用迭代优化方法对模型进行求解。利用Salinas和Pavia University高光谱影像进行对比实验,所提方法与分类方法结合用于影像分类时,其分类精度优于其他方法,总体分类精度分别达到97.42%和97.64%。     

Land use change detection based on multi‐date imagery from different satellite sensor systems

An empirical study was performed assessing the accuracy of land use change detection when using satellite image data acquired ten years apart by sensors with differing spatial resolutions. Landsat/Multi‐spectral Scanner (MSS) with Landsat/Thematic Mapper (TM) or SPOT/High Resolution Visible (HRV) multi‐spectral (XS) data were used as a multi‐data pair for detecting land use change. The primary objectives of the study were to: (1) compare standard change detection methods (e.g. multi‐date ratioing and principal components analysis) applied to image data of varying spatial resolution; (2) assess whether to transform the raster grid of the higher resolution image data to that of the lower resolution raster grid or vice‐versa in the registration process: and (3) determine if Landsat/TM or SPOT/ HRV(XS) data provides more accurate detection of land use changes when registered to historical Landsat/MSS data.

Ratioing multi‐sensor, multi‐date satellite image data produced higher change detection accuracies than did principal components analysis and is useful as a land use change enhancement technique. Ratioing red and near infrared bands of a Landsat/MSS‐SPOT/HRV(XS) multi‐date pair produced substantially higher change detection accuracies (～10%) than ratioing similar bands of a Landsat/MSS ‐ Landsat/TM multi‐data pair. Using a higher‐resolution raster grid of 20 meters when registering Landsat/MSS and SPOTZHRV(XS) images produced a slightly higher change detection accuracy than when both images were registered to an 80 meter raster grid. Applying a “majority”; moving window filter whose size approximated a minimum mapping unit of 1 hectare increased change detection accuracies by 1–3% and reduced commission errors by 10–25%.     

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.     

Identification of hazelnut fields using spectral and Gabor textural features

Land cover identification and monitoring agricultural resources using remote sensing imagery are of great significance for agricultural management and subsidies. Particularly, permanent crops are important in terms of economy (mainly rural development) and environmental protection. Permanent crops (including nut orchards) are extracted with very high resolution remote sensing imagery using visual interpretation or automated systems based on mainly textural features which reflect the regular plantation pattern of their orchards, since the spectral values of the nut orchards are usually close to the spectral values of other woody vegetation due to various reasons such as spectral mixing, slope, and shade. However, when the nut orchards are planted irregularly and densely at fields with high slope, textural delineation of these orchards from other woody vegetation becomes less relevant, posing a challenge for accurate automatic detection of these orchards. This study aims to overcome this challenge using a classification system based on multi-scale textural features together with spectral values. For this purpose, Black Sea region of Turkey, the region with the biggest hazelnut production in the world and the region which suffers most from this issue, is selected and two Quickbird archive images (June 2005 and September 2008) of the region are acquired. To differentiate hazel orchards from other woodlands, in addition to the pansharpened multispectral (4-band) bands of 2005 and 2008 imagery, multi-scale Gabor features are calculated from the panchromatic band of 2008 imagery at four scales and six orientations. One supervised classification method (maximum likelihood classifier, MLC) and one unsupervised method (self-organizing map, SOM) are used for classification based on spectral values, Gabor features and their combination. Both MLC and SOM achieve the highest performance (overall classification accuracies of 95% and 92%, and Kappa values of 0.93 and 0.88, respectively) when multi temporal spectral values and Gabor features are merged. High F_β values (a combined measure of producer and user accuracy) for detection of hazel orchards (0.97 for MLC and 0.94 for SOM) indicate the high quality of the classification results. When the classification is based on multi spectral values of 2008 imagery and Gabor features, similar F_β values (0.95 for MLC and 0.93 for SOM) are obtained, favoring the use of one imagery for cost/benefit efficiency. One main outcome is that despite its unsupervised nature, SOM achieves a classification performance very close to the performance of MLC, for detection of hazel orchards.