Hyperspectral Images-Based Crop Classification Scheme for Agricultural Remote Sensing

Hyperspectral imaging is gaining a significant role in agricultural remote sensing applications. Its data unit is the hyperspectral cube which holds spatial information in two dimensions while spectral band information of each pixel in the third dimension. The classification accuracy of hyperspectral images (HSI) increases significantly by employing both spatial and spectral features. For this work, the data was acquired using an airborne hyperspectral imager system which collected HSI in the visible and near-infrared (VNIR) range of 400 to 1000 nm wavelength within 180 spectral bands. The dataset is collected for nine different crops on agricultural land with a spectral resolution of 3.3 nm wavelength for each pixel. The data was cleaned from geometric distortions and stored with the class labels and annotations of global localization using the inertial navigation system. In this study, a unique pixel-based approach was designed to improve the crops' classification accuracy by using the edge-preserving features (EPF) and principal component analysis (PCA) in conjunction. The preliminary processing generated the high-dimensional EPF stack by applying the edge-preserving filters on acquired HSI. In the second step, this high dimensional stack was treated with the PCA for dimensionality reduction without losing significant spectral information. The resultant feature space (PCA-EPF) demonstrated enhanced class separability for improved crop classification with reduced dimensionality and computational cost. The support vector machines classifier was employed for multiclass classification of target crops using PCA-EPF. The classification performance evaluation was measured in terms of individual class accuracy, overall accuracy, average accuracy, and Cohen kappa factor. The proposed scheme achieved greater than 90 % results for all the performance evaluation metrics. The PCA-EPF proved to be an effective attribute for crop classification using hyperspectral imaging in the VNIR range. The proposed scheme is well-suited for practical applications of crops and landfill estimations using agricultural remote sensing methods.     

Sequential masking classification of multi‐temporal Landsat7 ETM+ images for field‐based crop mapping in Karacabey,Turkey

Three Landsat7 ETM+ images acquired in May, July and August during the 2000 crop growing season were used for field‐based mapping of summer crops in Karacabey, Turkey. First, the classification of each image date was performed on a standard per pixel basis. The results of per pixel classification were integrated with digital agricultural field boundaries and a crop type was determined for each field based on the modal class calculated within the field. The classification accuracy was computed by comparing the reference data, field‐by‐field, to each classified image. The individual crop accuracies were examined on each classified data and those crops whose accuracy exceeds a preset threshold level were determined. A sequential masking classification procedure was then performed using the three image dates, excluding after each classification the class properly classified. The final classified data were analysed on a field basis to assign each field a class label. An immediate update of the database was provided by directly entering the results of the analysis into the database. The sequential masking procedure for field‐based crop mapping improved the overall accuracies of the classifications of the July and August images alone by more than 10%.     

Study on Differences between Sentinel-2A and Landsat-8 Images in Rape Identification

Rape is one of the most important crops for many countries,so it is important to obtain accurate rape area.Compared with Landsat-8 data,Sentinel-2A has many advantages,but whether the results of Sentinel-2A data in crop identification are better than Landsat-8 is still an unknown question.The study site is located in a typical agricultural region：Gaochun District in Nanjing,the capital of Jiangsu Province,China,with central coordinates of 118°52′E and 31°19′N.One Sentinel-2A and one Landsat-8 image were obtained during the flowering stage of rape,and then rape area was extracted by using different classification methods based on spectral characteristics and vegetation indices.By comparing the identification accuracy of two images under different classification conditions and methods,the results show that：(1) The difference of spectral characteristics and separability of vegetation indices of different objects in Sentinel-2A were higher than those of Landsat-8 images;(2) Under the classifier of support vector machine,the Producer’s and User’s accuracy of rape of Sentinel-2A based on spectral characteristics were 89.7% and 91.3% respectively,which were 7.0% and 6.2% higher than the identification accuracy of Landsat-8 data;(3) After adding texture information,the overall accuracy and kappa coefficient of two kinds of data were significantly improved,but there was no increase in the producer’s and user’s accuracy of rape.The result presented in this paper show that compared with Landsat-8 data,Sentinel-2A data is more suitable for extracting crop distribution information in small areas with complex planting structure,which can lay a theoretical foundation for crop identification and application of Sentinel-2A data.     

Study on Crops Remote Sensing Classification based onMulti-temporal Landsat 8 OLI Images

It is crucial for agricultural production to know crop planting situation.Temporal remote sensing images and subtle spectral characteristics of ground features play an important role in extracting crops distribution.At this point,multi-temporal Landsat 8 OLI images were used to extracting the distribution of main crops in the east of Xinrong district of Datong city by using Spectral Angle Mapper(SAM) combined with the decision tree classification,and the extracting result was compared with the result that maximum likelihood extracted.The results show that：① The planting area of spring corn,grain,soybean and potato is decreased and mosaic distribution in order.② The overall accuracy obtained by SAM combined with the decision tree classification is 85.34% and the Kappa coefficient is 0.76,which is outperformed the results of maximum likelihood with the increase of 22.51% and 0.31,respectively,the classification results was more consistent with the actual distribution of main crops.③ The classification accuracy of main crops used the multi-temporal remote sensing images was obviously higher than that of single-temporal image,and the difference between ground features and spectra in middle or high resolution images can effectively weaken by analyzing multi-temporal data from the perspective of difference of spectral angle.The results not only confirmed the positive effect of multi-temporal remote sensing images on crops classification,but also developed the SAM combined with decision tree classification in crops classification of medium-high resolution remote sensing images,which has a certain application prospect.     

Hybrid Multi-Strategy Aquila Optimization with Deep Learning Driven Crop Type Classification on Hyperspectral Images

Hyperspectral imaging instruments could capture detailed spatial information and rich spectral signs of observed scenes. Much spatial information and spectral signatures of hyperspectral images (HSIs) present greater potential for detecting and classifying fine crops. The accurate classification of crop kinds utilizing hyperspectral remote sensing imaging (RSI) has become an indispensable application in the agricultural domain. It is significant for the prediction and growth monitoring of crop yields. Amongst the deep learning (DL) techniques, Convolution Neural Network (CNN) was the best method for classifying HSI for their incredible local contextual modeling ability, enabling spectral and spatial feature extraction. This article designs a Hybrid Multi-Strategy Aquila Optimization with a Deep Learning-Driven Crop Type Classification (HMAODL-CTC) algorithm on HSI. The proposed HMAODL-CTC model mainly intends to categorize different types of crops on HSI. To accomplish this, the presented HMAODL-CTC model initially carries out image preprocessing to improve image quality. In addition, the presented HMAODL-CTC model develops dilated convolutional neural network (CNN) for feature extraction. For hyperparameter tuning of the dilated CNN model, the HMAO algorithm is utilized. Eventually, the presented HMAODL-CTC model uses an extreme learning machine (ELM) model for crop type classification. A comprehensive set of simulations were performed to illustrate the enhanced performance of the presented HMAODL-CTC algorithm. Extensive comparison studies reported the improved performance of the presented HMAODL-CTC algorithm over other compared methods.     

基于CNN和农作物光谱纹理特征进行作物分布制图

以卷积神经网络（Convolutional Neural Network, CNN）为代表的深度学习技术,在农作物遥感分类制图领域具有广阔的应用前景。以多时相Landsat 8 多光谱遥感影像为数据源,搭建CNN模型对农作物进行光谱特征提取与分类,并与支撑向量机（SVM）常规分类方法进行对比。进一步引入影像纹理信息,利用CNN对农作物光谱和纹理特征进行提取,优化作物分布提取结果。实验表明：① 基于光谱特征的农作物分布提取,验证结果对比显示,CNN对应各类别精度、总体精度均优于SVM,其中二者总体精度分别为95.14%和91.77%;② 引入影像纹理信息后,基于光谱和纹理特征的CNN农作物分类总体精度提高至96.43%,Kappa系数0.952,且分类结果的空间分布更为合理,可有效区分花生、道路等精细地物,说明纹理特征可用于识别不同作物。基于光谱和纹理信息的CNN特征提取,可面向种植结构复杂区域实现农作物精准分类与分布制图。     

基于CNN和农作物光谱纹理特征进行作物分布制图

以卷积神经网络（Convolutional Neural Network, CNN）为代表的深度学习技术，在农作物遥感分类制图领域具有广阔的应用前景。以多时相Landsat 8 多光谱遥感影像为数据源，搭建CNN模型对农作物进行光谱特征提取与分类，并与支撑向量机（SVM）常规分类方法进行对比。进一步引入影像纹理信息，利用CNN对农作物光谱和纹理特征进行提取，优化作物分布提取结果。实验表明：① 基于光谱特征的农作物分布提取，验证结果对比显示，CNN对应各类别精度、总体精度均优于SVM，其中二者总体精度分别为95.14%和91.77%；② 引入影像纹理信息后，基于光谱和纹理特征的CNN农作物分类总体精度提高至96.43%，Kappa系数0.952，且分类结果的空间分布更为合理，可有效区分花生、道路等精细地物，说明纹理特征可用于识别不同作物。基于光谱和纹理信息的CNN特征提取，可面向种植结构复杂区域实现农作物精准分类与分布制图。     

Accurate crop-type classification using multi-temporal optical and multi-polarization SAR data in an object-based image analysis framework

Accurate crop-type classification is a challenging task due, primarily, to the high within-class spectral variations of individual crops during the growing season (phenological development) and, second, to the high between-class spectral similarity of crop types. Utilizing within-season multi-temporal optical and multi-polarization synthetic aperture radar (SAR) data, this study introduces a combined object- and pixel-based image classification methodology for accurate crop-type classification. Particularly, the study investigates the improvement of crop-type classification by using the least number of multi-temporal RapidEye (RE) images and multi-polarization Radarsat-2 (RS-2) data utilized in an object- and pixel-based image analysis framework. The method was tested on a study area in Manitoba, Canada, using three different classifiers including the standard Maximum Likelihood (ML), Decision Tree (DT), and Random Forest (RF) classifiers. Using only two RE images of July and August, the proposed method results in overall accuracies (OAs) of about 95%, 78%, and 93% for the ML, DT, and RF classifiers, respectively. Moreover, the use of only two quad-pol images of RS-2 of June and September resulted in OAs of 92%, 75%, and 90% for the ML, DT, and RF classifiers, respectively. The best classification results were achieved by the synergistic use of two RE and two RS-2 images. In this case, the overall classification accuracies were 97% for both ML and RF classifiers. In addition, the average producer’s accuracies of 95% and 96% were achieved by the ML and RF classifiers, respectively, whereas the average user accuracy was 94% for both classifiers. The results indicated promising potentials for rapid and cost-effective local-scale crop-type classification using a limited number of high-resolution optical and multi-polarization SAR images. Very accurate classification results can be considered as a replacement for sampling the agricultural fields at the local scale. The result of this very accurate classification at discrete locations (approximately 25 × 25 km frames) can be applied in a separate procedure to increase the accuracy of crop area estimation at the regional to provincial scale by linking these local very accurate spatially discrete results to national wall-to-wall continuous crop classification maps.     

Analysis of natural images processing for the extraction of agricultural elements

This work presents several developed computer-vision-based methods for the estimation of percentages of weed, crop and soil present in an image showing a region of interest of the crop field. The visual detection of weed, crop and soil is an arduous task due to physical similarities between weeds and crop and to the natural and therefore complex environments (with non-controlled illumination) encountered. The image processing was divided in three different stages at which each different agricultural element is extracted: (1) segmentation of vegetation against non-vegetation (soil), (2) crop row elimination (crop) and (3) weed extraction (weed). For each stage, different and interchangeable methods are proposed, each one using a series of input parameters which value can be changed for further refining the processing. A genetic algorithm was then used to find the best value of parameters and method combination for different sets of images. The whole system was tested on several images from different years and fields, resulting in an average correlation coefficient with real data (bio-mass) of 84%, with up to 96% correlation using the best methods on winter cereal images and of up to 84% on maize images. Moreover, the method’s low computational complexity leads to the possibility, as future work, of adapting them to real-time processing.     

Cover Airborne MSS data processing for forest classification

Abstract

Four SPOT HRV images of the same area of East Anglia, acquired between February and September 1986, have been evaluated at the National Remote Sensing Centre for their potential use in agricultural land cover mapping. Spectral coincidence plots were used in feature selection. Information from single images contained a high level of spectral confusion between cover types. Vegetation index images and original data were used in supervised maximum likelihood classification. Higher classification accuracies were achieved using the original data than the vegetation indices. An overall classification accuracy of 71 per cent for 10 land cover types was improved to 88 per cent by reducing the number of classes. Although the imagery acquired for the study did not correspond well to key dates in the crop calendar, the broad land cover categories, cereal crops, field crops (sugar beet and vegetables), grass land and broadleaved woodlands could be mapped from SPOT. Using vegetation indices from the whole scene, a map of land cover has been produced for an administrative district within the scene. Comparison with simulated Thematic Mapper data indicates greater crop discrimination is provided in the mid-infrared part of the spectrum.     

Spectral mixture analysis of agricultural crops: Endmember validation and biophysical estimation in potato plots

The spectral reflectance of agricultural crops is affected significantly by sub-pixel scale spectral contributions of background soils and shadows as viewed by a remote sensing instrument. This has meant the potential of remote sensing imagery has not been fully realized for extracting biophysical information and assessing ecological stress using methods such as vegetation indices (VIs). In this paper, we address this problem explicitly using spectral mixture analysis (SMA) to quantify the area abundance of plants, soils and shadows at sub-pixel scales with the aim of improving extraction of plant biophysical and structural information from remote sensing data. Different measurement strategies were tested in the field for acquiring reference endmember spectra of crop vegetation, soil and shadows using a field spectroradiometer for a set of potato plots in western Canada. Endmember measurements included sunlit and shadowed spectra of in situ crop targets, optically thick stacks and data from excised leaves, as well as cultivated, rough and compacted soils. All possible combinations of crop, soil and shadow endmember spectra were analysed using SMA to derive sets of sub-pixel scale component fractions from radiometer spectra acquired from a boom truck over replicate plot samples with a sensor field of view of 1.05 m. Digital video image frames captured simultaneously with the radiometer data were used to determine ground proportions of crop, soil and shadow for independent validation of the SMA fractions. Endmember fractions derived from excised leaves, cultivated soil and shadowed vegetation spectra showed the best agreement with ground truth data, with differences of only ±3.3%. These sub-pixel scale fractions were used in regression analyses to predict leaf area index, biomass and plant width with an average r2 value of 0.85 from SMA shadow fraction, which was a substantial improvement over the best VI results from NDVI, NGVI and SR (average r2 = 0.53). Perspectives on SMA at different stages in the growing season and for different crop types are provided with a recommendation that further SMA research is warranted for local to regional scale agricultural crop monitoring programmes.     

Integrating multiple texture methods and NDVI to the Random Forest classification algorithm to detect tea and hazelnut plantation areas in northeast Turkey

Hazelnuts and tea are two major agricultural crops grown in the eastern Black Sea region in Turkey. Since this part of Turkey is not industrialized, most of the local people work in agriculture, making hazelnuts and tea a part of their lives. For the government side, it is crucial to keep records of the amount of harvested croplands to implement agricultural policies. In fact, the harvested area and crop type of each cadastral parcel are collected either during cadastral surveys or with the declaration of individual farmers, yet this information is mostly not up-to-date and does not reflect the current land-use status. This study aims to determine the extent and distribution of hazelnuts and tea grown areas using the Random Forest (RF) classification algorithm. Tea and hazelnuts give similar spectral reflectance values to surrounding vegetation, which makes it difficult to distinguish them using only their spectral properties. To tackle this problem, the normalized difference vegetation index (NDVI) and texture extraction methods such as the Grey Level Co-occurrence Matrix (GLCM) and Gabor filter were integrated with the RF algorithm, and their contributions to the success of the RF classification method were examined. WorldView-2 satellite images, which have eight multispectral bands (MS: 2 m) and one higher spatial resolution panchromatic band (PAN: 0.5 m), were used. Since the study area contains agricultural products grown in different seasons, satellite images belonging to both summer and winter periods were used. Preliminary results acquired using only spectral values indicated that the RF method gives 79.05% and 71.84% overall accuracies for summer and winter periods, respectively. Integrating texture information improves the performance of the RF algorithm such that the overall classification accuracies are increased to 83.54% and 87.89% when texture information extracted with GLCM and the Gabor filter is added. The classification performance of the winter image is also boosted to be 77.41% and 79.73% with the contribution of texture information obtained with GLCM and the Gabor filter, respectively. Finally, produced thematic maps were compared with the latest cadastral maps to validate classification results with ground truth data. The obtained results reveal the success of integrating texture features in classification since the created thematic maps are consistent with actual land use. The results also show that the crops grown on some cadastral parcels are not coherent with the most current cadastral database, which implies that the cadastral maps need to be updated.     

Knowledge-based crop classification of a Landsat Thematic Mapper image

Abstract

A knowledge-based classification method was designed to improve crop classification accuracy. Crop data of preceding years, stored in a geographical information system (GIS) were used as ancillary data. Knowledge about crop succession, determined from crop rotation schemes, was formalized by means of transition matrices. The spectral data, the data from the GIS and the knowledge represented in the transition matrix were used in a modified Bayesian classification algorithm. The developed classification was tested in an agricultural region in The Netherlands. Depending on the spectral class discrimination, the accuracy of the knowledge-based classification was 6 to 20 percent better compared with a maximum likelihood classification.     

A computationally efficient multi-domain active learning method for crop mapping using satellite image time-series

Crop mapping through classification of Satellite Image Time-Series (SITS) data can provide precious information for several agricultural applications, such as crop monitoring, yield forecasting, and crop inventory. However, several issues affect the classification performance of SITS data. As one of the most challenging problems, constituent images of time-series provide different levels of information about crops. These differences are the result of dynamic spectral responses of crops and also the variable atmospheric and sensor conditions. The second issue is the unavailability of adequate high-quality samples for training the classifier. In this study, we proposed a novel computationally efficient Multi-Domain Active Learning (MDAL) method which takes advantage of Multiple Kernel Learning (MKL) and Active Learning (AL) algorithms to address these two issues. The proposed method uses MKL algorithms to address the issues associated with different information level of the data, which generally cannot be modelled using the well-known classification algorithms. AL algorithms were also used for semi-automatic selection of training samples. However, most of the MKL algorithms are very computationally demanding. Consequently, using them in the MDAL method can dramatically increase the computational costs. Thus, in this paper, we presented the similarity-based MKL algorithms. Thanks to their low computational complexities, these algorithms are the most suitable MKL algorithms that can be used in the MDAL method. We evaluated the proposed method using two multispectral SITS datasets, acquired by RapidEye and SPOT sensors. The obtained results of the MDAL method for these datasets respectively showed 8.2% and 5.87% increase in the overall accuracy of classification as compared to the accuracy of the standard AL algorithm. The results also showed that in the case of adopting the SimpleMKL algorithm (a common MKL algorithm in the literature) the computational time of the MDAL method is 577 and 474 seconds for RapidEye and SPOT datasets, respectively. However, in the case of adopting the similarity-based MKL algorithms, these computational times respectively decreases to 4 and 2 seconds.     

Assessing the spatio-temporal rates and patterns of land-use and land-cover changes in the Cerrados of southeastern Mato Grosso,Brazil

The Cerrados of central Brazil have undergone profound landscape transformation in recent decades due to agricultural expansion, and this remains poorly assessed. The present research investigates the spatial-temporal rates and patterns of land-use and land-cover (LULC) changes in one of the main areas of agricultural production in Mato Grosso State (Brazil), the region of Primavera do Leste. To quantify the different aspects of LULC changes (e.g. rates, types, and spatial patterns) in this region, we applied a post-classification change detection method, complemented with landscape metrics, for three dates (1985, 1995, and 2005). LULC maps were obtained from an object-based classification approach, using the nearest neighbour (NN) classifier and a multi-source data set for image object classification (e.g. seasonal Thematic Mapper (TM) bands, digital elevation model (DEM), and a Moderate Resolution Imaging Spectroradiometer (MODIS)-derived index), strategically chosen to increase class separability. The results provided an improved mapping of the Cerrados natural vegetation conversion into crops and pasture once auxiliary data were incorporated into the classification data set. Moreover, image segmentation was crucial for LULC map quality, in particular because of crop size and shape. The changes detected point towards increasing loss and fragmentation of natural vegetation and high rates of crop expansion. Between 1985 and 2005, approximately 42% (6491 km²) of Cerrados in the study area were converted to agricultural land uses. In addition, it was verified that cultivated areas are encroaching into fragile environments such as wetlands, which indicates the intense pressure of agricultural expansion on the environment.     

农作物品质遥感反演研究进展

当今农业生产管理迫切需要直接迅速的信息指导。随着科技水平的不断提高,通过利用不同遥感技术手段,实现实时监测农作物生长过程中的主要影响因子,使无损预测预报农作物品质成为可能。通过分析几种农作物的主要品质性状及形成影响因素,在归纳农作物品质监测常用光谱参量的基础上,从地面平台和航天航空平台两方面分别介绍近年来国内外主要研究进展,总结农作物品质遥感监测模型建立使用的主要算法,综合分析农作物品质遥感监测技术实现过程中存在的若干问题,同时提出相应的解决措施,并对遥感监测技术进行了展望。     

Object-based crop identification using multiple vegetation indices, textural features and crop phenology

Crop identification on specific parcels and the assessment of soil management practices are important for agro-ecological studies, greenhouse gas modeling, and agrarian policy development. Traditional pixel-based analysis of remotely sensed data results in inaccurate identification of some crops due to pixel heterogeneity, mixed pixels, spectral similarity, and crop pattern variability. These problems can be overcome using object-based image analysis (OBIA) techniques, which incorporate new spectral, textural and hierarchical features after segmentation of imagery. We combined OBIA and decision tree (DT) algorithms to develop a methodology, named Object-based Crop Identification and Mapping (OCIM), for a multi-seasonal assessment of a large number of crop types and field status.In our approach, we explored several vegetation indices (VIs) and textural features derived from visible, near-infrared and short-wave infrared (SWIR) bands of ASTER satellite scenes collected during three distinct growing-season periods (mid-spring, early-summer and late-summer). OCIM was developed for 13 major crops cultivated in the agricultural area of Yolo County in California, USA. The model design was built in four different scenarios (combinations of three or two periods) by using two independent training and validation datasets and the best DTs resulted in an error rate of 9% for the three-period model and between 12 and 16% for the two-period models. Next, the selected DT was used for the thematic classification of the entire cropland area and mapping was then evaluated applying the confusion matrix method to the independent testing dataset that reported 79% overall accuracy. OCIM detected intra-class variations in most crops attributed to variability from local crop calendars, tree-orchard structures and land management operations. Spectral variables (based on VIs) contributed around 90% to the models, although textural variables were necessary to discriminate between most of the permanent crop-fields (orchards, vineyard, alfalfa and meadow). Features extracted from late-summer imagery contributed around 60% in classification model development, whereas mid-spring and early-summer imagery contributed around 30 and 10%, respectively. The Normalized Difference Vegetation Index (NDVI) was used to identify the main groups of crops based on the presence and vigor of green vegetation within the fields, contributing around 50% to the models. In addition, other VIs based on SWIR bands were also crucial to crop identification because of their potential to detect field properties like moisture, vegetation vigor, non-photosynthetic vegetation and bare soil. The OCIM method was built using interpretable rules based on physical properties of the crops studied and it was successful for object-based feature selection and crop identification.     

广东特色农业波谱数据库设计与开发

自20世纪70年代以来,国际和国内波谱数据库的发展虽然如火如荼,但是都存在不同层次的缺陷,不能满足现阶段我国遥感基础研究和应用的需要,对华南特色农业遥感应用来说差距更远,该研究以建立波谱知识实用型库为目标,集成了华南农作物波谱、环境参数、应用模型,试建立了基于WEB的广东荔枝、龙眼、甘蔗等特色农作物波谱数据库,实现了特色农作物波谱数据库的概念设计、数据的组织、功能和界面设计等方面,为多样的华南特色农作物波谱数据库的建立提供了一个示范。     

基于视觉显著性的农作物图像评价方法研究

近年来,随着电子信息技术、计算机可视化技术和互联网技术的迅速发展,传统的农业管理方式正逐渐被新型的农业信息化管理方法所替代。当前,基于数字图像处理技术的智慧农业已经成为新型农业信息化中的关键研究领域。具体来说,利用数字平台获取的农作物图像可以为专家提供非常丰富的信息,比如,农作物的长势,农作物病虫害情况等。但是,值得注意的是,获取这些有价值信息的前提是数字平台拍摄的农作物图像具有足够的清晰度,即没有出现严重失真。基于此,以农作物图像的质量为研究对象,提出了一种基于视觉显著性的半参考质量评价方法,在梯度域提取了方向直方图特征来刻画图像质量的变化。实验结果表明,该方法能够很好地识别农作物的图像质量,保证了后续高层次信息提取的有效性。     

Selecting patterns and features for between- and within- crop-row weed mapping using UAV-imagery

This paper approaches the problem of weed mapping for precision agriculture, using imagery provided by Unmanned Aerial Vehicles (UAVs) from sunflower and maize crops. Precision agriculture referred to weed control is mainly based on the design of early post-emergence site-specific control treatments according to weed coverage, where one of the most important challenges is the spectral similarity of crop and weed pixels in early growth stages. Our work tackles this problem in the context of object-based image analysis (OBIA) by means of supervised machine learning methods combined with pattern and feature selection techniques, devising a strategy for alleviating the user intervention in the system while not compromising the accuracy. This work firstly proposes a method for choosing a set of training patterns via clustering techniques so as to consider a representative set of the whole field data spectrum for the classification method. Furthermore, a feature selection method is used to obtain the best discriminating features from a set of several statistics and measures of different nature. Results from this research show that the proposed method for pattern selection is suitable and leads to the construction of robust sets of data. The exploitation of different statistical, spatial and texture metrics represents a new avenue with huge potential for between and within crop-row weed mapping via UAV-imagery and shows good synergy when complemented with OBIA. Finally, there are some measures (specially those linked to vegetation indexes) that are of great influence for weed mapping in both sunflower and maize crops.