Airborne Hyperspectral Imagery and Yield Monitor Data for Mapping Cotton Yield Variability

Increased availability of hyperspectral imagery necessitates the evaluation of its potential for precision agriculture applications. This study examined airborne hyperspectral imagery for mapping cotton (Gossypium hirsutum L.) yield variability as compared with yield monitor data. Hyperspectral images were acquired using an airborne imaging system from two cotton fields during the 2001 growing season, and yield data were collected from the fields using a cotton yield monitor. The raw hyperspectral images contained 128 bands between 457 and 922 nm. The raw images were geometrically corrected, georeferenced and resampled to 1 m resolution, and then converted to reflectance. Aggregation functions were then applied to each of the 128 bands to reduce the cell resolution to 4 m (close to the cotton picker's cutting width) and 8 m. The yield data were also aggregated to the two grids. Correlation analysis showed that cotton yield was significantly related to the image data for all the bands except for a few bands in the transitional range from the red to the near-infrared region. Stepwise regression performed on the yield and hyperspectral data identified significant bands and band combinations for estimating yield variability for the two fields. Narrow band normalized difference vegetation indices derived from the significant bands provided better yield estimation than most of the individual bands. The stepwise regression models based on the significant narrow bands explained 61% and 69% of the variability in yield for the two fields, respectively. To demonstrate if narrow bands may be better for yield estimation than broad bands, the hyperspectral bands were aggregated into Landsat-7 ETM+ sensor's bandwidths. The stepwise regression models based on the four broad bands explained only 42% and 58% of the yield variability for the two fields, respectively. These results indicate that hyperspectral imagery may be a useful data source for mapping crop yield variability.     

Evaluating high resolution SPOT 5 satellite imagery to estimate crop yield

High resolution satellite imagery has the potential to map within-field variation in crop growth and yield. This study examined SPOT 5 satellite multispectral imagery for estimating grain sorghum yield. A 60 km × 60 km SPOT 5 scene and yield monitor data from three grain sorghum fields were recorded in south Texas. The satellite scene contained four spectral bands (green, red, near-infrared and mid-infrared) with a 10-m spatial resolution. Subsets were extracted from the scene that covered the three fields. Images with pixel sizes of 20 and 30 m were also generated from the individual field images to simulate coarser resolution satellite imagery. Vegetation indices and principal components were derived from the images at the three spatial resolutions. Grain yield was related to the vegetation indices, the four bands and the principal components for each field, and for all the fields combined. The effect of the mid-infrared band on estimates of yield was examined by comparing the regression results from all four bands with those from the other three bands. Statistical analysis showed that the 10-m, four-band image and the aggregated 20-m and 30-m images explained 68, 76 and 83%, respectively, of the variation in yield for all the fields combined. The coefficient of determination between yield and the imagery increased with pixel size because of the smoothing effect. The inclusion of the mid-infrared band slightly improved the R ² values. These results indicate that high resolution SPOT 5 multispectral imagery can be a useful data source for determining within-field yield variation for crop management.     

Comparison of QuickBird Satellite Imagery and Airborne Imagery for Mapping Grain Sorghum Yield Patterns

Timely and accurate information on crop conditions obtained during the growing season is of vital importance for crop management. High spatial resolution satellite imagery has the potential for mapping crop growth variability and identifying problem areas within fields. The objectives of this study were to use QuickBird satellite imagery for mapping plant growth and yield patterns within grain sorghum fields as compared with airborne multispectral image data. A QuickBird 2.8-m four-band image covering a cropping area in south Texas, USA was acquired in the 2003 growing season. Airborne three-band imagery with submeter resolution was also collected from two grain sorghum fields within the satellite scene. Yield monitor data collected from the two fields were resampled to match the resolutions of the airborne imagery and the satellite imagery. The airborne imagery was related to yield at original submeter, 2.8 and 8.4 m resolutions and the QuickBird imagery was related to yield at 2.8 and 8.4 m resolutions. The extracted QuickBird images for the two fields were then classified into multiple zones using unsupervised classification and mean yields among the zones were compared. Results showed that grain yield was significantly related to both types of image data and that the QuickBird imagery had similar correlations with grain yield as compared with the airborne imagery at the 2.8 and 8.4 m resolutions. Moreover, the unsupervised classification maps effectively differentiated grain production levels among the zones. These results indicate that high spatial resolution satellite imagery can be a useful data source for determining plant growth and yield patterns for within-field crop management.     

Mapping Grain Sorghum Yield Variability Using Airborne Digital Videography

Mapping crop yield variability is one important aspect of precision agriculture. Combine-mounted yield monitors are becoming widely available for measuring and mapping yields for different crops. This study was designed to assess airborne digital videography as a tool for mapping grain sorghum yields for precision farming. Color-infrared (CIR) imagery was acquired with a three-camera digital video imaging system from two grain sorghum fields in south Texas over the 1995 and 1996 growing seasons. The multispectral video data obtained during the bloom to soft dough stages of plant development were related to hand-harvested grain yields at sampling sites determined from unsupervised image classification maps of the two fields. Significant correlations were found between grain yields and the red band, the green band, and the normalized difference vegetation index (NDVI). Regression equations were developed to describe the relations between grain yields and each of the three significant spectral variables using an exponential model and two segmented models. Multiple linear regression equations were also determined to relate grain yields to the three bands and NDVI. These equations were then used to estimate grain yields at each video image pixel within each field and to generate grain yield maps. Comparisons of the estimated average yields from the regression equations with the actual yields indicated that yield estimation errors from the equations ranged from 0.0 to 10.0% in 1995 and from 0.2 to 7.3% in 1996 for field 1, and from 4.0 to 11.2% in 1995 and 6.3 to 12.5% in 1996 for field 2. Although the equations developed for one field in a given year may not apply to the same field in any other year, the practical value of these relationships is for mapping within-field grain yield variations. The results from this study showed that airborne digital videography, combined with ground sampling, regression analysis, and image processing, could be a useful approach for mapping spatial crop yield variability within fields.     

Airborne hyperspectral imagery and linear spectral unmixing for mapping variation in crop yield

Spectral unmixing techniques can be used to quantify crop canopy cover within each pixel of an image and have the potential for mapping the variation in crop yield. This study applied linear spectral unmixing to airborne hyperspectral imagery to estimate the variation in grain sorghum yield. Airborne hyperspectral imagery and yield monitor data recorded from two sorghum fields were used for this study. Both unconstrained and constrained linear spectral unmixing models were applied to the hyperspectral imagery with sorghum plants and bare soil as two endmembers. A pair of plant and soil spectra derived from each image and another pair of ground-measured plant and soil spectra were used as endmember spectra to generate unconstrained and constrained soil and plant cover fractions. Yield was positively related to the plant fraction and negatively related to the soil fraction. The effects of variation in endmember spectra on estimates of cover fractions and their correlations with yield were also examined. The unconstrained plant fraction had essentially the same correlations (r) with yield among all pairs of endmember spectra examined, whereas the unconstrained soil fraction and constrained plant and soil fractions had r-values that were sensitive to the spectra used. For comparison, all 5151 possible narrow-band normalized difference vegetation indices (NDVIs) were calculated from the 102-band images and related to yield. Results showed that the best plant and soil fractions provided better correlations than 96.3 and 99.9% of all the NDVIs for fields 1 and 2, respectively. Since the unconstrained plant fraction could represent yield variation better than most narrow-band NDVIs, it can be used as a relative yield map especially when yield data are not available. These results indicate that spectral unmixing applied to hyperspectral imagery can be a useful tool for mapping the variation in crop yield.     

Using spectral distance, spectral angle and plant abundance derived from hyperspectral imagery to characterize crop yield variation

Vegetation indices (VIs) derived from remote sensing imagery are commonly used to quantify crop growth and yield variations. As hyperspectral imagery is becoming more available, the number of possible VIs that can be calculated is overwhelmingly large. The objectives of this study were to examine spectral distance, spectral angle and plant abundance (crop fractional cover estimated with spectral unmixing) derived from all the bands in hyperspectral imagery and compare them with eight widely used two-band and three-band VIs based on selected wavelengths for quantifying crop yield variability. Airborne 102-band hyperspectral images acquired at the peak development stage and yield monitor data collected from two grain sorghum fields were used. A total of 64 VI images were generated based on the eight VIs and selected wavelengths for each field in this study. Two spectral distance images, two spectral angle images and two abundance images were also created based on a pair of pure plant and soil reference spectra for each field. Correlation analysis with yield showed that the eight VIs with the selected wavelengths had r values of 0.73–0.79 for field 1 and 0.82–0.86 for field 2. Although all VIs provided similar correlations with yield, the modified soil-adjusted vegetation index (MSAVI) produced more consistent r values (0.77–0.79 for field 1 and 0.85–0.86 for field 2) among the selected bands. Spectral distance, spectral angle and plant abundance produced similar r values (0.76–0.78 for field 1 and 0.83–0.85 for field 2) to the best VIs. The results from this study suggest that either a VI (MSAVI) image based on one near-infrared band (800 or 825 nm) and one visible band (550 or 670 nm) or a plant abundance image based on a pair of pure plant and soil spectra can be used to estimate relative yield variation from a hyperspectral image.     

Utility of Remote Sensing in Predicting Crop and Soil Characteristics

Remote sensing during the production season can provide visual indications of crop growth along with the geographic locations of those areas. A grid coordinate system was used to sample cotton and soybean fields to determine the relationship between spectral radiance, soil parameters, and cotton and soybean yield. During the 2 years of this study, mid- to late-season correlation coefficients between spectral radiance and yield generally ranged from 0.52 to 0.87. These correlation coefficients were obtained using the green–red ratio and a vegetation index similar to the normalized difference vegetation index (NDVI) using the green and red bands. After 102 days after planting (DAP), the ratio vegetation index (RVI), difference vegetation index (DVI), NDVI, and soil-adjusted vegetation index (SAVI) generally provided correlation coefficients from 0.54 to 0.87. Correlation coefficients for cotton plant height measurements taken 57 and 66 DAP during 2000 ranged from 0.51 to 0.76 for all bands, ratios, and indices examined, with the exception of Band 4 (720nm). The most consistent correlation coefficients for soybean yield were obtained 89–93 DAP, corresponding to peak vegetative production and early pod set, using RVI, DVI, NDVI, and SAVI. Correlation coefficients generally ranged from 0.52 to 0.86. When the topographic features and soil nutrient data were analyzed using principal component analysis (PCA), the interaction between the crop canopy, topographic features, and soil parameters captured in the imagery allowed the formation of predictive models, indicating soil factors were influencing crop growth and could be observed by the imagery. The optimum time during 1999 and 2000 for explaining the largest amount of variability for cotton growth occurred during the period from first bloom to first open boll, with R values ranging from 0.28 to 0.70. When the PCA-stepwise regression analysis was performed on the soybean fields, R ² values were obtained ranging from 0.43 to 0.82, 15 DAP, and ranged from 0.27 to 0.78, 55–130 DAP. The use of individual bands located in the green, red, and NIR, ratios such as RVI and DVI, indices such as NDVI, and stepwise regression procedures performed on the cotton and soybean fields performed well during the cotton and soybean production season, though none of these single bands, ratios, or indices was consistent in the ability to correlate well with crop and soil characteristics over multiple dates within a production season. More research needs to be conducted to determine whether a certain image analysis method will be needed on a field-by-field basis, or whether multiple analysis procedures will need to be performed for each imagery date in order to provide reliable estimates of crop and soil characteristics.     

Predicting Cotton Lint Yield Maps from Aerial Photographs

It is generally accepted that aerial images of growing crops provide spatial and temporal information about crop growth conditions and may even be indicative of crop yield. The focus of this study was to develop a straightforward technique for creating predictive cotton yield maps from aerial images. A total of ten fields in southern Georgia, USA, were studied during three growing seasons. Conventional (true color) aerial photographs of the fields were acquired during the growing season in two to four week intervals. The aerial photos were then digitized and analyzed using an unsupervised classification function of image analysis software. During harvest, conventional yield maps were created for each of the fields using a cotton picker mounted yield monitor. Classified images and yield maps were compared quantitatively and qualitatively. A pixel by pixel comparison of the classified images and yield maps showed that spatial agreement between the two gradually increased in the weeks after planting, maintained spatial agreement of between 40% and 60% during weeks eight to fourteen, and then gradually declined again. The highest spatial agreement between a classified image and a yield map was 78%. The highest average agreement was 52% and occurred 9.9 weeks after planting. The visual similarity between the classified images and the yield maps were striking. In all cases, the dates with the best visual agreement occurred between eight and ten weeks after planting, and generally, during July for southern Georgia. This method offers great potential for offering cotton farmers early-season maps that predict the spatial distribution of yield. Although these maps can not provide magnitudes, they clearly show the resulting yield patterns. With inherent knowledge of past performance, farmers can use this information to allocate resources, address crop growth problems, and, perhaps, improve the profitability of their farm operation. These maps are well suited to be offered to farmers as a service by a crop consultant or a cooperative.     

Mapping crop ground cover using airborne multispectral digital imagery

Empirical relationships between remotely sensed vegetation indices and canopy density information, such as leaf area index or ground cover (GC), are commonly used to derive spatial information in many precision farming operations. In this study, we modified an existing methodology that does not depend on empirical relationships and extended it to derive crop GC from high resolution aerial imagery. Using this procedure, GC is calculated for every pixel in the aerial imagery by dividing the perpendicular vegetation index (PVI) of each pixel by the PVI of full canopy. The study was conducted during the summer growing seasons of 2007 and 2008, and involves airborne and ground truth data from 13 agricultural fields in the Southern High Plains of the USA. The results show that the method described in this study can be used to estimate crop GC from high-resolution aerial images with an overall accuracy within 3% of their true values.     

Strategies for monitoring within-field soybean yield using Sentinel-2 Vis-NIR-SWIR spectral bands and machine learning regression methods

Soybean crop plays an important role in world food production and food security, and agricultural production should be increased accordingly to meet the global food demand. Satellite remote sensing data is considered a promising proxy for monitoring and predicting yield. This research aimed to evaluate strategies for monitoring within-field soybean yield using Sentinel-2 visible, near-infrared and shortwave infrared (Vis/NIR/SWIR) spectral bands and partial least squares regression (PLSR) and support vector regression (SVR) methods. Soybean yield maps (over 500 ha) were recorded by a combine harvester with a yield monitor in 15 fields (3 farms) in Paraná State, southern Brazil. Sentinel-2 images (spectral bands and 8 vegetation indices) across a cropping season were correlated to soybean yield. Information pooled across the cropping season presented better results compared to single images, with best performance of Vis/NIR/SWIR spectral bands under PLSR and SVR. At the grain filling stage, field-, farm- and global-based models were evaluated and presented similar trends compared to leaf-based hyperspectral reflectance collected at the Brazilian National Soybean Research Center. SVR outperformed PLSR, with a strong correlation between observed and predicted yield. For within-field soybean yield mapping, field-based SVR models (developed individually for each field) presented the highest accuracies. The results obtained demonstrate the possibility of developing within-field yield prediction models using Sentinel-2 Vis/NIR/SWIR bands through machine learning methods.

     

Airborne Videography to Identify Spatial Plant Growth Variability for Grain Sorghum

Much research has focused on the use of intensive grid soil sampling and yield monitors to identify within-field spatial variability in precision farming. This paper reports on the use of airborne videography to identify spatial plant growth patterns for grain sorghum. Color-infrared (CIR) digital video images were acquired from two grain sorghum fields in south Texas several times during the 1995 and 1996 growing seasons. The video images were registered, and classified into several zones of homogeneous spectral response using an unsupervised classification procedure. Ground truthing was performed upon a limited number of sites within each zone to determine plant density, plant height, leaf area index, biomass, and grain yield. Results from both years showed that the digital video imagery identified within-field plant growth variability and that classification maps effectively differentiated grain production levels and growth conditions within the two fields. A temporal comparison of the images and classification maps indicated that plant growth patterns differed somewhat between the two successive growing seasons, though areas exhibiting consistently high or low yield were identified within each field.     

Unmanned aerial vehicle canopy reflectance data detects potassium deficiency and green peach aphid susceptibility in canola

There is growing evidence that potassium deficiency in crop plants increases their susceptibility to herbivorous arthropods. The ability to remotely detect potassium deficiency in plants would be advantageous in targeting arthropod sampling and spatially optimizing potassium fertilizer to reduce yield loss due to the arthropod infestations. Four potassium fertilizer regimes were established in field plots of canola, with soil and plant nutrient concentrations tested on three occasions: 69 (seedling), 96 (stem elongation), and 113 (early flowering) days after sowing (DAS). On these dates, unmanned aerial vehicle (UAV) multi-spectral images of each plot were acquired at 15 and 120 m above ground achieving spatial (pixel) resolutions of 8.1 and 65 mm, respectively. At 69 and 96 DAS, field plants were transported to a laboratory with controlled lighting and imaged with a 240-band (390–890 nm) hyperspectral camera. At 113 DAS, all plots had become naturally infested with green peach aphids (Hemiptera: Aphididae), and intensive aphid counts were conducted. Potassium deficiency caused significant: (1) increase in concentrations of nitrogen in youngest mature leaves, (2) increase in green peach aphid density, (3) decrease in vegetation cover, (4) decrease in normalized difference vegetation indices (NDVI) and decrease in canola seed yield. UAV imagery with 65 mm spatial resolution showed higher classification accuracy (72–100 %) than airborne imagery with 8 mm resolution (69–94 %), and bench top hyperspectral imagery acquired from field plants in laboratory conditions (78–88 %). When non-leaf pixels were removed from the UAV data, classification accuracies increased for 8 mm and 65 mm resolution images acquired 96 and 113 DAS. The study supports findings that UAV-acquired imagery has potential to identify regions containing nutrient deficiency and likely increased arthropod performance.     

Monitoring nitrogen status of potatoes using small unmanned aerial vehicles

Small unmanned aircraft vehicles (UAV) are potential remote sensing platforms for precision agriculture. However, to be useful for in-season management, nitrogen status needs to be estimated sufficiently early in the growing season. To determine when differences in nitrogen status of irrigated potatoes could be detected, an experiment was established in 2013 with a randomized block design with four N fertilization rates and three replicates. Over the growing season, a small parafoil-wing UAV was used to acquire color-infrared images with pixel sizes between 20 and 25 mm. Two normalized difference spectral indices were determined from image digital numbers, the normalized difference vegetation index (NDVI) and the green normalized difference vegetation index (GNDVI), which were then calibrated using reflectance-based NDVI and GNDVI. Unexpectedly, there were decreases in the NDVI and GNDVI calibrations with increased camera exposure time. After calibration, both NDVI and GNDVI were about equal to indices calculated using reflectances from high-altitude aerial photography and the WorldView-2 satellite. During tuber initiation and early tuber bulking, differences in measured leaf area index (LAI), chlorophyll meter values and spectral indices were only detectable at the lowest N fertilization rate. Later in the growing season, all N treatments could be distinguished in the imagery, but too late to mitigate yield losses from N deficiency. Linear relationships between plot GNDVI and NDVI were hypothesized to differ among N treatments because there would be less chlorophyll content per leaf area. Contrary to the hypothesis, there were no differences among fertilization rates on either of the two sampling dates. Compared with alternative technologies, small UAV platforms and sensors may not provide value to farmers for in-season nitrogen management.     

Maximizing the relationship of yield to site-specific management zones with object-oriented segmentation of hyperspectral images

Quick and low cost delineation of site-specific management zones (SSMZ) would improve applications of precision agriculture. In this study, a new method for delineating SSMZ using object-oriented segmentation of airborne imagery was demonstrated. Three remote sensing domains—spectral, spatial, and temporal- are exploited to improve the SSMZ relationship to yield. Common vegetation indices (VI), and first and second derivatives (\(\rho^{\prime}\), \(\rho^{\prime\prime}\)) from twelve airborne hyperspectral images of a cotton field for one season \(\rho^{\prime}\) were used as input layers for object-oriented segmentation. The optimal combination of VI, SSMZ size and crop phenological stage were used as input variables for SSMZ delineation, determined by maximizing the correlation to segmented yield monitor maps. Combining narrow band vegetation indices and object-oriented segmentation provided higher correlation between VI and yield at SSMZ scale than that at pixel scale by reducing multi-resource data noise. VI performance varied during the cotton growing season, providing better SSMZ delineation at the beginning and middle of the season (days after planting (DAP) 66–143).The optimal scale determined for SSMZ delineation was approximately 240 polygons for the study field, but the method also provided flexibility enabling the setting of practical scales for a given field. For a defined scale, the optimal single phenological stage for the study field was near July 11 (DAP 87) early in the growing season. SSMZs determined from multispectral VIs at a single stage were also satisfactory; compared to hyperspectral indices, temporal resolution of multi-spectral data seems more important for SSMZ delineation.     

基于植被指数对望奎县粮食作物产量预测方法的研究

以望奎县为研究区,利用TM影像确定望奎县土地利用现状,根据当年农作物生长季节内累积植被指数的平均值之和与农作物产量存在的关系,建立了最大植被指数遥感估产模型。并使用2007～2010年的MODIS遥感资料和该估产模型,对2007～2010年的望奎县粮食总产进行了估测,其中,前2个年度的估产误差都接近3%。此方法在大范围粮食总产预测中值得借鉴。     

Monthly composites from Sentinel-1 and Sentinel-2 images for regional major crop mapping with Google Earth Engine

Rapid and accurate access to large-scale, high-resolution crop-type distribution maps is important for agricultural management and sustainable agricultural development. Due to the limitations of remote sensing image quality and data processing capabilities, large-scale crop classification is still challenging. This study aimed to map the distribution of crops in Heilongjiang Province using Google Earth Engine(GEE) and Sentinel-1 and Sentinel-2 images. We obtained Sentinel-1 and Sentinel-2 images from all the covered study areas in the critical period for crop growth in 2018(May to September), combined monthly composite images of reflectance bands, vegetation indices and polarization bands as input features, and then performed crop classification using a Random Forest(RF) classifier. The results show that the Sentinel-1 and Sentinel-2 monthly composite images combined with the RF classifier can accurately generate the crop distribution map of the study area, and the overall accuracy(OA) reached 89.75%. Through experiments, we also found that the classification performance using time-series images is significantly better than that using single-period images. Compared with the use of traditional bands only(i.e., the visible and near-infrared bands), the addition of shortwave infrared bands can improve the accuracy of crop classification most significantly, followed by the addition of red-edge bands. Adding common vegetation indices and Sentinel-1 data to the crop classification improved the overall classification accuracy and the OA by 0.2 and 0.6%, respectively, compared to using only the Sentinel-2 reflectance bands. The analysis of timeliness revealed that when the July image is available, the increase in the accuracy of crop classification is the highest. When the Sentinel-1 and Sentinel-2 images for May, June, and July are available, an OA greater than 80% can be achieved. The results of this study are applicable to large-scale, high-resolution crop classification and provide key technologies for remote sensing-based crop classification in small-scale agricultural areas.     

An automated stand-alone in-field remote sensing system (SIRSS) for in-season crop monitoring

A stand-alone in field remote sensing system (SIRSS) with high spatial and temporal resolution was developed in this study. System control and image processing algorithms consisted of image acquisition control, camera parameter control, crop canopy reflectance calibration, image rectification, image background segmentation and vegetation indices map generation were developed and embedded in the SIRSS. The SIRSS is able to automatically capture multispectral images over a testing field at any predefined time points during the growing season and process captured images in real-time. This paper presents the SIRSS system design, image analysis procedures and determination of vegetation indices. In a validation experiment over an 8-plot corn field with three different nutrient treatments spanning the 2006 growing season, a total of 91 images were acquired and four different vegetation indices were derived from the images of each day. The largest differences of indices values among three treatments were indentified during the V6-V8 stages which implied this period could be the best time to detect variability caused by the nitrogen stress in the cornfield. The SIRSS has shown the potential of monitoring changes in vegetation status and condition.     

基于穗光谱指数的水稻产量预测

为探索运用水稻穗光谱植被指数预测水稻产量的可行性,以2个水稻品种为材料,设置3个氮素水平,测定了3个时期水稻叶片和穗的高光谱反射(350~2 500 nm)和色素含量,并测定了水稻的产量构成组分和籽粒产量。结果表明:与典型的植物反射光谱相比,水稻穗的反射光谱具有“绿峰消失”的特征;与叶片光谱指数[归一化差值指数(normalized vegetation index,NDVI)和光化学反射指数(photochemical reflectance index,PRI)]相比,穗光谱指数对叶绿素更敏感,而且能更准确地区分氮素水平。水稻叶片NDVI和PRI预测产量的均方根误差(RSME)分别为873.4~1 125.0、723.3~889.4 kg·hm^-2,而穗NDVI和PRI预测产量的RSME分别为681.7~743.1、515.0~637.8 kg·hm^-2,表明水稻穗光谱指数比叶片光谱指数更适合于水稻产量预测。