Sreekala G. Bajwa

HYPERSPECTRAL IMAGE DATA MINING FOR BAND SELECTION IN AGRICULTURAL APPLICATIONS

Hyperspectral remote sensing produces large volumes of data, quite often requiring hundreds of megabytes to gigabytes of memory storage for a small geographical area for one-time data collection. Although the high spectral resolution of hyperspectral data is quite useful for capturing and discriminating subtle differences in geospatial characteristics of the target, it contains redundant information at the band level. The objective of this study was to identify those bands that contain the most information needed for characterizing a specific geospatial feature with minimal redundancy. Band selection is performed with both unsupervised and supervised approaches. Five methods (three unsupervised and two supervised) are proposed and compared to identify hyperspectral image bands to characterize soil electrical conductivity and canopy coverage in agricultural fields. The unsupervised approach includes information entropy measure and first and second derivatives along the spectral axis. The supervised approach selects hyperspectral bands based on supplemental ground truth data using principal component analysis (PCA) and artificial neural network (ANN) based models. Each hyperspectral image band was ranked using all five methods. Twenty best bands were selected by each method with the focus on soil and plant canopy characterization in precision agriculture. The results showed that each of these methods may be appropriate for different applications. The entropy measure and PCA were quite useful for selecting bands with the most information content, while derivative methods could be used for identifying absorption features. ANN measure was the most useful in selecting bands specific to a target characteristic with minimum information redundancy. The results also indicated that a combination of wavebands with different bandwidths will allow use of fewer than 20 bands used in this study to represent the information contained in the top 20 bands, thus reducing image data dimensionality and volume considerably.

Aerial CIR remote sensing for weed density mapping in a soybean field

Accurate weed maps are essential for the success of site–specific herbicide application using map–based variable–rate sprayers. In this study, remotely sensed images acquired using an airborne digital color infrared (CIR) sensor were used for mapping and modeling the spatial distribution of weed infestation density within a soybean field. The effect of spatial positioning error associated with data on resolution requirements and mapping accuracy was also studied. Vegetative indices developed from the three–band CIR image showed strong correlation with spatial weed density. The best correlation was observed at the spatial resolutions of 4.5 m/pixel to 5.3 m/pixel, which was lower than the actual data resolutions. Higher modeling accuracies observed at lower resolutions were caused by the positioning error associated with both aerial imaging data and ground–truth data. At this resolution, the weed density models developed using an artificial neural network resulted in R 2 values of 0.87 and 0.83. This model mapped the spatial distribution of weed density with an R 2 value of 0.58 for a field not used in modeling.

SOIL FERTILITY CHARACTERIZATION IN AGRICULTURAL FIELDS USING HYPERSPECTRAL REMOTE SENSING

Airborne hyperspectral images provide high spatial and spectral resolution along with flexible temporal resolution that are ideally suited for precision agricultural applications. In this study, we have explored the potential of aerial visible/infrared (VIR) hyperspectral imagery for characterizing soil fertility factors in midwestern agricultural fields. Two fields (SW and NW) in Illinois and two fields (GV and FO) in Missouri were considered in this study. Field data included hyperspectral VIR images and soil fertility parameters including pH, organic matter (OM), Ca, Mg, P, K, and soil electrical conductivity. The VIR images were geo-registered and calibrated into apparent reflectance values. The FO field had the highest average reflectance, followed by SW, GV, and NW. The Illinois fields (SW and NW) were high in soil minerals, OM, and soil electrical conductivity. The measured soil fertility characteristics were modeled on first derivatives of the reflectance data using partial least square regression (PLSR). The PLSR model on derivative spectra was able to explain 66% of the overall variability in soil fertility variables considered in this study, with a predicted residual sum of square (PRESS) of 0.66. The model explained a higher degree of variability in some of the response variables, such as Ca (82%), Mg (72%), Veris shallow (86%), Veris deep (67%), and OM (66%), compared to factors such as pH (48%) and EM (50%). Analysis of the parameter estimates for each response variable showed that some of the wavebands, such as 625, 652, 658, 661, 754 and 784 nm, explained a high degree of variability in the model, whereas a large number of wavelengths had negligible contribution. In conclusion, this study showed that soil fertility factors important for precision agriculture applications can be successfully modeled on hyperspectral VIR remote sensing data with partial least square regression models.

Optimal Substitution of Cotton Burr and Linters in Thermoplastic Composites

A study was conducted to evaluate various substitutions of cotton burr and linters from cotton gin waste (CGW) as natural fiber reinforcements in ligno-cellulosic polymer composites (LCPC). Samples were fabricated with approximately 50 percent natural fiber, 40 percent high-density polyethylene, 4 percent mineral filler, and 6 percent lubricant, by weight. The experiment included substituting wood fiber in LCPC with 25, 50, 75, and 100 percent (by weight) cotton burr (CB) and cotton burr mixed with 2 percent (by weight) second-cut linters (CBL), respectively, with the remaining fraction as wood fiber and comparing it against the control (100% wood). Samples were extruded into rectangular profiles and tested for physical and mechanical properties such as specific gravity (SG), water absorption, thickness swelling, coefficients of linear thermal expansion (CLTE), flexural strength and modulus, compressive strength, hardness, and nail withdrawal force (NWF). The CB and CBL treatments exhibited SG, CLTE, hardness, and NWF comparable to the control. However, the water absorption and thickness swelling, flexural strength and modulus, and compressive strength all deteriorated at high substitution rates of CB and CBL. The favorable properties of cotton burr included its tendency to decrease CLTE and increase hardness of LCPC.

Effect of Laboratory Aging on the Physical and Mechanical Properties of Wood-Polymer Composites

The long-term performance of wood-polymer composites (WPC) under severe weather conditions is not well known. This study evaluates the changes in physical and mechanical properties of three commercially available WPC and treated southern yellow pine (SYP) under a modified 6-cycle accelerated aging process. The accelerated aging causes warping, splitting, discoloration, and significant changes in physical and mechanical properties of SYP. The compressive and flexural strength of the WPCs show negligible changes whereas stiffness, hardness, and screw withdrawal force show considerable deterioration and some recovery during accelerated aging. The composition and manufacturing process influence the performance of WPC under accelerated aging.

MODELING RICE PLANT NITROGEN EFFECT ON CANOPY REFLECTANCE WITH PARTIAL LEAST SQUARE REGRESSION (PLSR)

In Arkansas and other major rice growing states in the U.S., nitrogen fertilizer is commonly applied in a split application. The split application consists of a pre-flood ground application and one or two mid-season applications. Currently there are no fast and accurate methods for estimating mid-season nitrogen requirements of a rice crop. The goal of this research was to test spectral characteristics of a rice canopy to model and predict nitrogen status of a rice crop at the beginning of internodal elongation (BIE). The early part of internodal elongation is a critical period for mid-season nitrogen application due to the increased demand for nitrogen in the reproductive stage. Plot experiments were conducted with two popular rice varieties, Cocodrie and Wells, and six nitrogen levels of 0, 34, 67, 101, 134, 168 kg N/ha (0, 30, 60, 90, 120, and 150 lbs/acre) in a completely randomized block design with four replications. Data on canopy reflectance, biomass, and tissue nitrogen were measured three times at 7-day intervals starting at BIE. A t-test comparison showed no significant difference between the two cultivars in plant nitrogen, tissue nitrogen, or biomass at BIE. Both tissue nitrogen and plant nitrogen showed a parabolic relationship with yield, with yield maximized at 13 g/m2 plant nitrogen and 3% tissue nitrogen. A partial least square regression (PLSR) model on canopy reflectance could explain 47%, 63%, and 71% of the variation in plant nitrogen for weeks 0, 1, and 2, respectively, from BIE. The cultivar did not affect the sensitivity of the PLSR models. With these models, it is possible to identify rice nitrogen requirements during the internodal elongation with moderate to high accuracy. A nitrogen management regime can be developed based on rice canopy reflectance using the PLSR models. The amount of nitrogen required by the field can be calculated as the difference between the predicted plant nitrogen status and the optimum level of 13 g/m2, at which the rice yield was maximized.

Investigation of the Effects of Soil Compaction in Cotton

Soil compaction can cause yield reductions in cotton (Gossypium hirsutum). This study investigated the effect of soil compaction on canopy spectral reflectance, soil electrical conductivity (EC), and cotton yield. Field experiments were conducted during 2003-2005 using a completely randomized block design with four soil compaction treatments. The treatments were no subsoiling (control); subsoiled, disked, and bedded (conventional); subsoiled and compacted (compaction I); and compacted with no subsoiling (compaction II). Field data were collected on soil resistance, canopy reflectance, soil EC, and cotton yield. Comparison of means showed differences between treatments in reflectance in 2003 and 2004, soil compaction parameters in 2004 and 2005, and soil EC and yield in 2005. The depth and thickness of the hardpan were significantly correlated to green NDVI on 16 September (R2 = 0.53) in 2003. Depth, average resistance of hard pan, and EC all showed relationships with yield in 2005. Their combination as independent variables could explain 65% of the variability in cotton yield in 2005. These results verified that compaction affected canopy reflectance and reduced cotton yield in Arkansas. The practical implications of the outcome of this study are the potential use of EC and canopy reflectance to infer crop yield and extent of soil compaction. However, a multi-site and multi-year study is necessary to confirm this possibility.

Spatial Correlation of Crop Response to Soybean Cyst Nematode (Heterodera glycines)

Heterodera glycines, the soybean cyst nematode (SCN), is a serious threat to U.S. soybean production and is difficult to detect at onset, as its symptoms are not dramatic. The objective of this study was to use spatial regression tools to map the spatial variability of spectral response of a soybean crop to SCN infestation. Field experiments were conducted in 2003 and 2004 on a soybean field at the University of Arkansas Pine Tree Experiment Station. The SCN distribution in the field was assessed twice each year. Airborne visible-infrared imagery was acquired during both seasons. Linear regression analysis showed that SCN had no effect (R2 < 0.2) on yield and green normalized difference vegetation index (GNDVI) derived from aerial imagery on a single date. Spatial regression models on SCN incorporating the semivariogram of residuals of the ordinary least square regression accounted for 70% of the variability in GNDVI and 78% of the variability in yield. They indicated that spatial regression is a better tool than simple linear regression for analyzing soybean response to SCN infestation.

Spectral Response of Cotton Canopy to Water Stress

Accurate irrigation scheduling is important to ensure maximum yield and optimal water use in irrigated cotton. However, irrigation scheduling research in the past has shown mixed yield responses to irrigation, which could be attributed to inaccuracies in the irrigation scheduling programs. This study hypothesizes that plant response to water stress, if monitored closely, could be a valuable parameter in irrigation scheduling. Field experiments were conducted in the 2003-2004 crop seasons with three different irrigation levels to study cotton response to water stress. The responses of canopy reflectance and temperature to water stress were analyzed. Although canopy temperature is a good indicator of water stress, it is directly related to vapor pressure deficits (VPD). Therefore, we also studied the effect of VPD on canopy temperature. Rainfall was plentiful in both seasons and high levels of water stress did not develop, resulting in no significant differences in lint yield associated with the irrigation treatments. Even in relatively wet years, both canopy temperature and reflectance showed great potential to indicate water stress. The reflectance-based vegetative indices, NDVI and GNDVI and canopy-temperature-based indices, ST and CWSI showed significant differences among irrigation treatments. These four indices were also highly correlated to soil moisture tension at 0.2 m depth. The results also verified that very large VPD (> 2.5 kPa) and extremely low VPD (< 1 kPa) masked the canopy temperature difference with respect to ambient temperature.

Soybean Disease Monitoring with Leaf Reflectance

Crop disease detection with remote sensing is a challenging area that can have significant economic and environmental impact on crop disease management. Spectroscopic remote sensing in the visible and near-infrared (NIR) region has the potential to detect crop changes due to diseases. Soybean cyst nematode (SCN) and sudden death syndrome (SDS) are two common soybean diseases that are extremely difficult to detect in the early stages under mild to moderate infestation levels. The objective of this research study was to relate leaf reflectance to disease conditions and to identify wavebands that best discriminated these crop diseases. A microplot experiment was conducted. Data collected included 800 leaf spectra, corresponding leaf chlorophyll content and disease rating of four soybean cultivars grown under different disease conditions. Disease conditions were created by introducing four disease treatments of control (no disease), SCN, SDS, and SCN+SDS. Crop data were collected on a weekly basis over a 10-week period, starting from 71 days after planting (DAP). The correlation between disease rating and selected vegetation indices (VI) were evaluated. Wavebands with the most disease discrimination capability were identified with stepwise linear discriminant analysis (LDA), logistic discriminant analysis (LgDA) and linear correlation analysis of pooled data. The identified band combinations were used to develop a classification function to identify plant disease condition. The best correlation (>0.8) between disease rating and VI occurred during 112 DAP. Both LDA and LgDA identified several bands in the NIR, red, green and blue regions as critical for disease discrimination. The discriminant models were able to detect over 80% of the healthy plants accurately under cross-validation but showed poor accuracy in discriminating individual diseases. A two-class discriminant model was able to identify 97% of the healthy plants and 58% of the infested plants as having some disease from the plant spectra.