Agricultural production in terms of economic and environmental demand requires increasingly efficient utilization of resources. Excessive use of nutrients may cause leaching, whereas deficits could lead to impediments in tapping full yield potential. Due to heterogeneity of fields, small-scale application of fertilizer provides means to encounter challenges that could arise and to improve resource efficiency.
As part of an ongoing research project, we have investigated the ability to increase nitrogen efficiency for winter wheat with two data mining approaches, an Artificial Neural Network (ANN) and a Support-Vector-Machine (SVM), for small-scale nitrogen application carried out in a field trial. Results obtained by either approach were further compared to outcomes of uniform treatment (UT). Each outcome was evaluated in accordance with crop yield and Nitrogen Cost-free Revenue (NCfR). Prior to in-field application, it was necessary to train both data mining models (ANN, SVM) in order to predict small-scale yield for each of the three split applications at particular points in time (SA1, SA2, SA3). Thus, varying input parameters accounting for heterogeneous conditions of a field were included in spatial data tuples. This way, a training dataset was set up which was used within subsequent data mining algorithms. These were carried out separately for each split application. Input parameters used to train both algorithms were chosen according to parameter availability at the time of each application. For the first split application, historical yield maps, maps of soil apparent electrical conductivity (ECa), and varying amounts of applied nitrogen at SA1 were used. Additionally, for the second split application a small-scale canopy spectral measurement of Red Edge Inflection Point (REIP) at SA2 and varying amounts of applied nitrogen at SA2 were used (additional parameters were included in the third split application).
At the time of applying the ANN (SVM) in-field, parameter sets pertaining to each spatially distinct unit were passed to the trained algorithms. Parameter selection within each set corresponded to that of the training and differed between SA1, SA2, and SA3. For each split application, the optimal amount of nitrogen needed to be determined at the outset. Hence, the algorithms, iteratively, estimated a set of various crop yields for every possible applicable nitrogen amount for each spatially distinct unit. From those nitrogen amounts and their corresponding crop yields, combinations were found optimal which resulted in the highest NCfR value (considering one split application at a time).
In regards to outcomes, an analysis of variances (ANOVA) shows no statistically significant yield depression for ANN and SVM modeling strategies when compared to UT. Average yield for each of the strategies ranges from 9.6 – 10.2 t/ha. SVM and ANN show a monetary benefit of 28 EUR/ha and 65 EUR/ha, compared to UT. Thus, nitrogen efficiency is improved considerably by about 10% (SVM) and 34% (ANN). Considering that spatial dependencies are not covered, results obtained from this test may be biased according to spatial influences.
Thus, we attempted to mitigate such drawbacks by applying a two-step procedure based on linear models provided by the SAS proc-mixed routine. This procedure ensures that small-scale autocorrelation and large-scale trends (e.g. soil quality) were considered. Consequently, statistically significant crop yield differences are observed between ANN and UT, with a yield depression of about 0.6 t/ha for the ANN strategy. In contrast to these findings, no statistically significant yield differences occur between SVM and UT. The average cumulated amount of nitrogen applied for each strategy ranges considerably between 130 kg N/ha (ANN) and 199 kg N/ha (UT). Considering the ANN-based strategy, fewer amounts of nitrogen, however, result in a lower crop yield. In contrast, using the SVM approach fewer nitrogen applied (179 kg N/ha) still delivered the same level of crop yield (as compared to UT). Differences in NCfR for each of the strategies show that only the SVM occurs to have a monetary advantage (13 EUR/ha compared to UT). Further comparison with the UT strategy shows improved nitrogen efficiency by about 10% (30%) with SVM (ANN).
Based on this evidence we, therefore, have reason to conclude that data mining tools are suitable for further optimizing the application of nitrogen. In our case, inputs needed for winter wheat production could be reduced to varying degrees. As of the current stage of this field trial, it has become apparent that economic evaluations are most reliable when using actually applied nitrogen and actually harvested crop yield. Part of our study involved geo-statistical adjustments of yield. Since this (simulated) variable was put in relation to actual applied amounts of fertilizer - instead of also using actual obtained yield - inaccuracies in estimating financial benefits may have occurred but have not been further investigated at this point. As nitrogen efficiency has been increased to a considerable extent, the use of data mining tools shall be emphasized nonetheless. In effect, less leaching could be expected.