Hyper-spectral remote sensing for weed and nitrogen stress detection

This study investigated the possibility of using data, acquired from airborne multi-spectral or hyper-spectral sensors, to detect nitrogen status and presence of weeds in crops; with the ultimate aim of contributing towards the development of a decision support system for precision crop management (PCM).A 24-waveband (spectrum range 475 to 910 nm) multi-spectral sensor was used to detect weeds in corn (Zea mays L.) and soybean ( Glycine max (L.) Merr.) in 1999. Analysis of variance (ANOVA), followed by Scheffe's test, were used to determine which wavebands displayed significant differences in aerial spectral data due to weed treatments. It was found that the radiance values were mainly indicative of the contribution of weeds to the total vegetation cover in various plots, rather than indicative of changes in radiance of the crops themselves, or of differences in radiance between the weed populations and the crop species.In the year 2000, a 72-waveband (spectrum range 407 to 949 nm) hyperspectral sensor was used to detect weeds in corn gown at three nitrogen levels (60, 120 and 250 kg N/ha). The weed treatments were: no control of weeds, control of grasses, control of broadleaved weeds and control of all weeds. Imagery was acquired at the early growth, tassel, and fully-mature stages of corn. Hyper-spectral measurements were also taken with a 512-waveband field spectroradiometer (spectrum range 270 to 1072 nm). Measurements were also carried out on crop physiological and associated parameters. ANOVA and contrast analyses indicated that there were significant (alpha = 0.05) differences in reflectance at certain wavebands, due to weed control strategies and nitrogen application rates. Weed controls were best distinguished at tassel stage. Nitrogen levels were most closely related to reflectance, at 498 nm and 671 nm, in the aerial data set. Differences in other wavebands, whether related to nitrogen or weeds, appeared to be dependent on the growth stage. Better results were obtained from aerial than ground-based spectral data.Regression models, representing crop biophysical parameters and yield in terms of reflectance, at one or more wavebands, were developed using the maximum r2 criterion. The coefficients of determination (r 2) were generally greater than 0.7 when models were based on spectral data obtained at the tassel stage. Models based on normalized difference vegetation indices (NDVI) were more reliable at estimating the validation data sets than were the reflectance models. The wavebands at 701 nm and 839 nm were the most prevalent in these models.Decision trees, artificial neural networks (ANNs), and seven other classifiers were used to classify spectral data into the weed and nitrogen treatment categories. Success rates for validation data were lower than 68% (mediocre) when training was done for all treatment categories, but good to excellent (up to 99% success) for classification into levels of one or the other treatment (i.e. weed or nitrogen) and also classification into pairs of levels within one treatment. Not one classifier was determined best for all situations.The results of the study suggested that spectral data acquired from airborne platforms can provide vital information on weed presence and nitrogen levels in cornfields, which might then be used effectively in the development of PCM systems