Agrivoltaic systems (AV) combine agricultural activities with the production of electricity from photovoltaic (PV) panels on the same land area. The concept of AV systems was introduced in 1982 by Goetzberger and Zastrow, but only more recently have the increased environmental concerns and the favorable economic and political frameworks stimulated a growing interest in this technology. A critical issue in the development of AV is the selection of crops that can grow profitably under the micrometeorological conditions generated by AV systems. This experiment studied the effect of four different shade depth treatments (AV1 = 27%, AV2 = 16%, AV3 = 9%, and AV4 = 18%) on the morphology, physiology, and yield of a soybean crop grown under a large-scale AV system. The field results were used to validate the output of a simulation platform that couples the crop model GECROS with a set of algorithms for the estimation and spatialisation of the shading, radiation, and crop-related outputs. Crop height, leaf area index (LAI), and specific leaf area (SLA) all increased under the most shaded AV areas compared to the full light (FL, control) conditions. On average, under an AV system, the grain yield and the number of pods per plant were reduced by 8% and 13%, and in only one area (AV2) was a slight increase in grain yield (+4.4%) observed in comparison to the FL. The normalised root mean square error (nRMSE) value of the predicted grain yield differed from the observed grain values of 12.9% for the FL conditions, 15.7% in AV1, 16.5% in AV2, 6.71% in AV3, and 2.82% in AV4. Although the model simulated the yield satisfactorily, the results of the RMSE revealed that the model tends to underestimate the yield with an increase in shade, particularly for the AV1 and AV2 conditions.

Estimating key crop parameters (e.g., phenology, yield prediction) is a prerequisite for optimizing agrifood supply chains through the use of satellite imagery, but requires timely and accurate crop mapping. The moment in the season and the number of training sites used are two main drivers of crop classification performance. The combined effect of these two parameters was analysed for tomato crop classification, through 125 experiments, using the three main machine learning (ML) classifiers (neural network, random forest, and support vector machine) using a response surface methodology (RSM). Crop classification performance between minority (tomato) and majority (‘other crops’) classes was assessed through two evaluation metrics: Overall Accuracy (OA) and G-Mean (GM), which were calculated on large independent test sets (over 400,000 fields). RSM results demonstrated that lead time and the interaction between the number of majority and minority classes were the two most important drivers for crop classification performance for all three ML classifiers. The results demonstrate the feasibility of preharvest classification of tomato with high performance, and that an RSM-based approach enables the identification of simultaneous effects of several factors on classification performance. SVM achieved the best grading performances across the three ML classifiers, according to both evaluation metrics. SVM reached highest accuracy (0.95 of OA and 0.97 of GM) earlier in the season (low lead time) and with less training sites than the other two classifiers, permitting a reduction in cost and time for ground truth collection through field campaigns.

Near-real-time, high-spatial-resolution leaf area index (LAI) maps would enable producers to monitor crop health and growth status, improving agricultural practices such as fertiliser and water management. LAI retrieval methods are numerous and can be divided into statistical and physically based methods. While statistical methods are generally subject to high site-specificity but possess high ease of implementation and use, physically based methods are more transferable, albeit more complex to use in operational settings. In addition, statistical methods need a large amount of data for calibration and subsequent validation, and this is only seldom feasible. Techniques based on predictive equations (PE_physical) represent a viable alternative, allowing the partial combination of statistical and physical methods merits while minimising their shortcomings. In this paper, predictive equation-based techniques were compared with four other methods: two radiative transfer model (RTM) inversion methods, one based on neural network (NNET) and one based on a look-up table (LUT), and two empirical methods (one using empirical models based on vegetation indices and in situ data and one based on empirical models found in the scientific literature). The methods were chosen based on common use. To evaluate the performance of the studied methods, the coefficient of determination (R²), root mean square error (RMSE), and normalised root mean square error (nRMSE, %) between the estimates and in situ LAI measurements were reported. The best PE_physical results, achieved by the OSAVI index (RMSE = 0.84 m² m⁻²), provided better performance for LAI recovery than the NNET-based RTM inversions (0.86 m² m⁻²) or the estimates made by LUT (0.94 m² m⁻²). Furthermore, the best PE_physical produced accuracies comparable to the best empirical model (RMSE = 0.71 m² m⁻²), calibrated through in situ data, and similar to the best literature model (RMSE = 0.76 m² m⁻²). These results indicated that PE_physical can be used to recover LAI with transferability comparable to literature models.