Application of remote sensing to agricultural field trials

Summary

Remote sensing techniques enable quantitative information about a field trial to be obtained instantaneously and non-destructively. The aim of this study was to identify a method that can reduce inaccuracies in field trial analysis, and to identify how remote sensing can support and/or replace conventional field measurements in field trials.

In the literature there is a certain consensus that the best bands from which characteristic spectral information about vegetation can be extracted are those in the visible (green and red) and infrared regions of the electromagnetic spectrum. This was confirmed in the present study by an analysis of multispectral scanner data ('Daedalus scanner') from field trials with cereals. The optimal bands that were thereby selected for explaining grain yield mostly contained the channels 5 (550- 600 nm), 7 (650-700 nm), and 9 (800-890 nm).

Multispectral aerial photography was found to be most appropriate for recording extensive field trials in a short period. In the present study, recordings were carried out with a single-engine aircraft, using two Hasselblad cameras for obtaining vertical photographs on black and white 70-mm aerial films. In this way, costs stayed within acceptable limits. The recording scale chosen, given the dimensions of the trials at the experimental farm of the Wageningen Agricultural University, where the research was carried out, was 1:8 000. Photographs were taken approximately fortnightly to keep in step with conventional field sampling. The film/filter combinations selected for obtaining a high spectral resolution and for matching bands 5, 7 and 9 of the Daedalus scanner, resulted in the following passbands:
green : 555-580 nm;
red :665-700 nm;
infrared :840-900 nm,

The densities of the objects on the film were measured by means of an automated Macbeth TD-504 densitometer. An aperture with a diameter of 0.25 mm was selected for the densitometer, in order to obtain a high spatial resolution at the scale of 1:8000, applicable to field trials with plots 3 metres wide. The measured densities were converted into exposure values, corrected for light falloff, and then a linear function was applied to convert them into reflectance factors. In this linear function the exposure time, relative aperture, transmittance of the optical system, irradiance, path radiance and atmospheric attenuation were incorporated. Reference targets with known reflectance characteristics were set up in the field during missions and recorded at the same camera setting and under the same atmospheric conditions as the field trials, in order to ascertain the parameters of the linear function.

Information about crop reflectance obtained from the literature suggested that reflectances in the visible region of the electromagnetic spectrum (green or red) would be most suitable for estimating soil cover, whereas reflectances in the infrared might be most suitable for estimating leaf area index (LAI). Other plant characteristics, such as dry matter weight or yield, may be estimated indirectly from reflectances. Field trials with cereals analysed during the present study showed that treatment effects shown by green and red reflectances tended to be opposite to those shown by LAI. Treatment effects shown by infrared reflectance tended to be similar to those shown by LAI, even at large LAI (6-8). The treatment effects manifest in reflectances were more stable in time than those for LAI. Coefficients of variation of residuals resulting from analyses of variance were systematically smaller for reflectances than for the LAI in all experiments: those for the infrared reflectance were particularly small. In general, critical levels in testing for treatment effects were smaller for the infrared reflectance than for the LAI, which indicates that the power for infrared reflectance was larger than for LAI.

Soil moisture content is not constant during the growing season and differences in soil moisture content greatly influence soil reflectance. Since a multitemporal analysis of remote sensing data was required, a correction had to be made for soil background when ascertaining the relationship between reflectances and crop characteristics. In the literature no index or reflectance model stood out as being suitable for estimating crop characteristics in agricultural field trials. Thus, in this monograph an appropriate simplified reflectance model is presented for estimating soil cover and LAI for green vegetation. First of all, soil cover is redefined as: the vertical projection of green vegetation and the relative area of shadows included, seen by a sensor pointing vertically downwards, relative to the total soil area (in this definition soil cover depends on the position of the sun). Then, the simplified reflectance model is based on the expression of the measured reflectance as a composite reflectance of plants and soil: the measured reflectance in the various passbands is a linear combination of soil cover and its complement, with the reflectances of the plants and of the soil as coefficients, respectively.

By using this model, it should, theoretically, be possible to correct for soil background when estimating soil cover by combination of measurements in the green and red passbands. In practice, however, all the procedures derived yielded poor results because the difference between green and red reflectances was so small. Thus, attention was focussed on estimating LAI.

For estimating LAI a corrected infrared reflectance was calculated by subtracting the contribution of the soil from the measured reflectance. Theoretically, combining the reflectance measurements obtained in the green, red and infrared passbands, enables the corrected infrared reflectance to be calculated, without knowing soil reflectances. The main assumption was that there is a constant ratio between the reflectances of bare soil in different passbands, independent of soil moisture content: this assumption is valid for many soil types. For the soil type at the experimental farm of the Agricultural University, the corrected infrared reflectance can be approximated by the difference between total measured infrared and red reflectances. Subsequently this corrected infrared reflectance was used for estimating LAI according to the inverse of a special case of the Mitscherlich function. This function contains two parameters that have to be ascertained empirically. Model simulations with the SAIL model (introduced by Verhoef, 1984) confirmed the potential of this simplified, semi-empirical, reflectance model for estimating LAI.

Analogous derivations were applied for a generative canopy (cereals) with yellowing leaves.

The estimation of LAI by reflectances yielded good results for the field trials with cereals analysed in this study. The presence of treatment effects could be shown with larger power and the coefficients of variation were smaller for this estimated LAI than for the one measured in the field. Regression curves of LAI on corrected infrared reflectance differed significantly in different trials with the same crop, particularly for the generative stage. This may have been caused by large systematic discrepancies between LAI measurements obtained with the conventional sampling techniques for two field trials, because of subjectivity in separating green from yellow leaves. To date, the best approach is to ascertain regression curves of LAI on corrected infrared reflectance for each field trial by incorporating a few additional plots, in which both the LAI and the reflectances are measured.