Proximal soil sensors and geostatistical tools in precision agriculture applications

Recognition of spatial variability is very important in precision agriculture applications. The use of proximal soil sensors and geostatistical techniques is highly recommended worldwide to detect spatial variation not only in fields but also within-field (micro-scale). This study involves, as a first step, the use of visible and near infrared (vis-NIR) spectroscopy to estimate soil key properties (6) and obtain high resolution maps that allow us to model the spatial variability in the soil. Different calibration models were developed using partial least square regression (PLSR) for different soil properties. These calibration models were evaluated by both cross-validation and independent validation. Results show good to excellent calibration models for most of soil properties under study in both cross-validation and independent validation. The on-line maps created using the collected on-line spectra and the calibration models previously estimated for each soil property were compared with three different maps (measured, predicted, error).The second step uses multivariate geostatistical analysis to develop three different geostatistical models (soil, spectral, fusion). The soil model includes 8 soil properties, spectral model includes 4 soil properties and the fusion model includes 12 soil properties. The three models were evaluated by cross-validation and the results show that the goodness of fitting can be considered as satisfactory for the soil model, whereas the performance of the spectral model was quite poor. Regarding the fusion model, it performed quite well, though the model generally underestimated the high values and overestimated the low values. An independent validation data set was used to evaluate the performance of the three models calculating three statistics: mean error (ME), as an indicator of bias; mean standardized squared error (MSSE), as an indicator of accuracy, and root mean squared error (RMSE), as an indicator of precision of estimation. Synthetically, the two, soil and fusion, models performed quite similarly, whereas the performance of the spectral model was much poorer. With regard to delineation of management zones (MZs), the factor cokriging analysis was applied using the three different models. The first factor (F1) for the soil and fusion models was related to soil properties that affect soil fertility, whereas for the spectral model was related to P (-0.88) and pH (-0.42).Based on the first factor of the soil and fusion models, three management zones were delineated and classified as low, medium and high fertility zones using isofrequency classes. Spatial similarity between the yield map and delineated MZs maps based on F1 for the soil and fusion models was calculated.The overall accordance between the two maps was 40.0 % for the soil model and 38.6 % for the fusion model. The two models performed quite similarly. These results can be interpreted as more than 50% of the yield variation was ascribable to more dynamic factors than soil parameters not included in this study, such as agro-meteorological conditions, plant diseases, nutrition stresses, etc. However, the results are quite promising for the application of the proposed approach in site-specific management.