Precision Agriculture最新文献

Canopy Chlorophyll Content (CCC) is an important physiological indicator that reflects the growth stage of trees. Accurate estimation of CCC facilitates dynamic monitoring and efficient forest management. In this study, we used high-resolution remote sensing images obtained by uncrewed aerial vehicles (UAVs) equipped with multispectral sensors (red, green, blue, near-infrared, and red-edge) to estimate CCC of lodgepole pine (Pinus elliottii). Our aim was to determine the optimal machine learning model between support vector regression (SVR) and random forest regression (RFR) for predicting CCC and to evaluate the effectiveness of multispectral bands along with 21 vegetation indices (VIs) in the estimation process. Individual tree boundaries were derived from the canopy height model (CHM) based on three-dimensional (3D) point clouds generated using structure from motion. These images, combined with continuous field measurements from January to December, provided comprehensive data for our analysis. The results showed that the SVR method outperformed the RFR method in estimating leaf chlorophyll content (LCC), with fitting R² values up to 0.692 and RMSE values up to 0.168 mg⋅g⁻¹. Overall, the study highlights the potential of UAV-based remote sensing for multitemporal forest monitoring, offering advances in precision forestry and tree breeding.

On-farm experiments are increasingly being used as their costs have decreased with technological advances in collecting, storing, and processing geospatial data. A question that has not been well addressed is what spatial experimental design is best for on-farm experiments when the goal is to estimate a spatially varying coefficients (SVC) model. The focus here is determining the optimal location of treatments to obtain a nearly D-optimal experimental design when estimating a linear plateau model. A pseudo-Bayesian approach is taken here because the field’s site-specific optimal nitrogen value is unknown. Optimal designs are generated, assuming a fixed number of replications for each treatment level. The resulting designs are more efficient than classic Latin square, strip plot, and completely randomized designs. The method consistently produces designs that have 95% efficiency or higher. Random designs had efficiencies varying from 41 to 64% with Latin squares having higher efficiencies and strip plots lower.

Abstract

Site-specific nitrogen management has been proposed as a tool to increase crop yield while decreasing nutrient losses to the environment. Many reports can be found on sensing technologies to quantify the variability within a field and the definition of management zones based on the observed variability. However, fewer studies have been dedicated to the selection of the most suitable N fertilizer management scenario: should more or less nutrients be applied in the zones with a lower crop productivity potential? To address this knowledge gap, nine Flemish maize fields were selected as potential candidates for precision fertilization based on the soil maps and historical vegetation index patterns. Within each field, two management zones were identified based on historical vegetation index patterns and electrical conductivity maps, and different fertilization strategies were tested in each zone. The field trial results in terms of yield and soil residual nitrate showed that site-specific N management outperforms the conventional practice only in the fields with temporally stable management zones. In the fields having differences in the physical soil properties (e.g. presence of stones or clay particles), affecting water availability, lower fertilization in zones with a poor soil productivity potential could be recommended. In the fields where the performance of the management zones changes from year to year mainly due to annual variation in precipitation, a risk of incorrect implementation of the precision fertilization concept was identified. Historical NDVI time series serve a good basis to delineate the temporally stable management zones.

In developing countries like Iran, where information is scarce, understanding the spatial variability of soil available phosphorous (SAP), one of the three major nutrients, is crucial for effective agricultural ecosystem management. This study aimed to predict and digitally map the spatial distribution and related uncertainty of SAP while also assessing the impact of environmental factors on SAP variability in the topsoils. A study area from northern Khuzestan province, Iran was selected as case study area. Three machine learning (ML) models, namely, Random Forest (RF), Artificial Neural Network (ANN), and Support Vector Regression (SVR), were used to develop predictive relationship between surface soil (0–10 cm) SAP content and environmental covariates derived from a digital elevation model and Landsat 8 images. A total of 250 topsoil samples were collected following the conditioned Latin Hypercube Sampling (cLHS) approach and several soil properties were measured in the laboratory. Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Lin’s Concordance Correlation Coefficient (LCCC) were used to determine the accuracy of models. The findings indicated that the RF algorithm demonstrated the most favorable performance, with a mean absolute error (MAE) of 0.85 mg SAP kg⁻¹, the lowest root mean square error (RMSE) of 0.99 mg SAP kg⁻¹, and the highest linear correlation coefficient (LCCC) values of 0.96. This suggests that the RF algorithm had the least tendency to overestimate or underestimate SAP contents compared to other methods. Consequently, the RF algorithm was selected as the optimal choice. Predictive ML models were employed to digitally map SAP contents within the region. Spatial patterns of SAP contents showed an increasing gradient from west to east. The spatial variability information provides a basis for developing sustainable production system in the area.

Pre-harvest yield prediction of direct-seeded rice is critical for guiding crop interventions and food security assessment in precision agriculture. Technology advances in unmanned aerial vehicle (UAV)-based remote sensing has provided an unprecedented opportunity to efficiently retrieve crop growth parameters instead of labor-intensive ground measurements. This study is aiming to evaluate the feasibility of fusing multi-temporal UAV-derived features collected at critical phenological stages in forecasting direct-seeded rice yield across different cultivars and nitrogen (N) management. The results showed that RGB sensor-derived canopy volume, canopy coverage, and spectral features including RBRI, WI etc., were identified to be most sensitive to the differences in aboveground biomass and grain yield. Heading stage was the suitable time for estimating yield performance (R² = 0.75) for mono-temporal UAV observation. By contrast, multi-temporal features fusion could remarkably enhance the yield prediction accuracy. Moreover, the yield prediction accuracy can be further improved by integrating UAV features collected at panicle initiation and heading stages (i.e., rice reproductive phase) compared to multi-temporal features fusion (R² increased from 0.82 to 0.85 and RMSE decreased from 35.1 to 31.5 g m⁻²). This can be attributed to the fact that the biomass accumulation during the reproductive phase was closely associated to the total spikelets and final yield. By using this proposed approach, the predicted yield showed a good spatial consistency with the measured yield across different cultivars and N management, and yield prediction error in the most of the plots (114 of 128 plots) was less than 45 g m⁻². In summary, this study highlights that the reproductive phase is the optimal time window for UAV observing, which provides an effective method for accurate pre-harvest yield prediction of direct-seeded rice in precision agriculture.

Crop yield prediction from UAV images has significant potential in accelerating and revolutionizing crop breeding pipelines. Although convolutional neural networks (CNN) provide easy, accurate and efficient solutions over traditional machine learning models in computer vision applications, a CNN training requires large number of ground truth data, which is often difficult to collect in the agricultural context. The major objective of this study was to develope an end-to-end 3D CNN model for plot-scale soybean yield prediction using multitemporal UAV-based RGB images with approximately 30,000 sample plots. A low-cost UAV-RGB system was utilized and multitemporal images from 13 different experimental fields were collected at Argentina in 2021. Three commonly used 2D CNN architectures (i.e., VGG, ResNet and DenseNet) were transformed into 3D variants to incorporate the temporal data as the third dimension. Additionally, multiple spatiotemporal resolutions were considered as data input and the CNN architectures were trained with different combinations of input shapes. The results reveal that: (a) DenseNet provided the most efficient result (R² 0.69) in terms of accuracy and model complexity, followed by VGG (R² 0.70) and ResNet (R² 0.65); (b) Finer spatiotemporal resolution did not necessarily improve the model performance but increased the model complexity, while the coarser resolution achieved comparable results; and (c) DenseNet showed lower clustering patterns in its prediction maps compared to the other models. This study clearly identifies that multitemporal observation with UAV-based RGB images provides enough information for the 3D CNN architectures to accurately estimate soybean yield non-destructively and efficiently.

In the natural environment, the detection and recognition process of Punna navel orange fruit using machine vision systems is affected by many factors, such as complex background, uneven light illumination, occlusions of branches and leaves and large variations in fruit size. To solve these problems of low accuracy in fruit detection and poor robustness of the detection algorithm in the field conditions, a new object detection algorithm, named Retinanet_G2S, was proposed in this paper based on the modified Retinanet network. The images of Punna navel orange were collected with Microsoft Kinect V2 in the uncontrolled environment. Firstly, a new Res2Net-GF network was designed to replace the section of feature extraction in the original Retinanet, which can potentially improve the learning ability of target features of the trunk network. Secondly, a multi-scale cross-regional feature fusion grids network was designed to replace the feature pyramid network module in the original Retinanet, which could enhance the ability of feature information fusion among different scales of the feature pyramid. Finally, the original border regression localization method in Retinanet network was optimized based on the accurate boundary box regression algorithm. The study results showed that, compared with the original Retinanet network, Retinanet_G2S improved mAP, mAP50, mAP75, mAP_S, mAP_M and mAP_L by 3.8%, 1.7%, 5.8%, 2.4%, 2.1% and 5.5%, respectively. Moreover, compared with 7 types of classic object detection models, including SSD, YOLOv3, CenterNet, CornerNet, FCOS, Faster-RCNN and Retinanet, the average increase in mAP of Retinanet_G2S was 9.11%. Overall, Retinanet_G2S showed a promising optimization effect, particularly for the detection of small targets and overlapping fruits.

An orchard sprayer prototype running a variable-rate algorithm to adapt the spray volume to the canopy characteristics (dimensions, shape and leaf density) in real-time was designed and implemented. The developed machine was able to modify the application rate by using an algorithm based on the tree row volume, in combination with a newly coefficient defined as Density Factor (Df). Variations in the canopy characteristics along the row crop were electronically measured using six ultrasonic sensors (three per sprayer side). These differences in foliage structure were used to adjust the flow rate of the nozzles by merging the ultrasonic sensors data and the forward speed information received from the on-board GNSS. A set of motor-valves was used to regulate the final amount of sprayed liquid. Laboratory and field tests using artificial canopy were arranged to calibrate and select the optimal ultrasonic sensor configuration (width beam and signal pre-processing method) that best described the physical canopy properties. Results indicated that the sensor setup with a medium beam width offered the most appropriate characterization of trees in terms of width and Df. The experimental sprayer was also able to calculate the application rate automatically depending on changes on target trees. In general, the motor valves demonstrated adequate capability to supply and control the required liquid pressure at all times, mainly when spraying in a range between 4.0 and 14.0 MPa. Further work is required on the equipment, such as designing field efficiency tests for the sprayer or refining the accuracy of Df.

Plant height, a key agronomic trait, affects crop structure, photosynthesis, and thus the final yield and seed quality. The combination of digital cameras on unmanned aerial vehicles (UAVs) and use of structure from motion have enabled high-throughput crop canopy height estimation. However, the focus of prior research has mainly been on plot-level height prediction, neglecting precise estimations for individual plants. This study aims to explore the potential of UAV RGB images with mask region-based convolutional neural network (Mask-RCNN) for high-throughput phenotyping of individual-level height (IH) in oilseed rape at different growth stages. Field-measured height (FH) of nine sampling plants in each subplot of the 150 subplots was obtained by manual measurement after the UAV flight. An instance segmentation model for oilseed rape with data augmentation based on the Mask-RCNN model was developed. The IHs were then used to obtain plot-level height based on individual-level height (PHIH). The results show that Mask-RCNN performed better than the conventional Otsu method with the F1 score increased by 60.8% and 26.6% under high and low weed pressure, respectively. The trained model with data augmentation achieved accurate crop height estimation based on overexposed and underexposed UAV images, indicating the model’s applicability in practical scenarios. The PHIH can be predicted with the determination coefficient (r²) of 0.992, root mean square error (RMSE) of 4.03 cm, relative root mean square error (rRMSE) of 7.68%, which outperformed the results in the reported studies, especially in the late bolting stage. The IHs of the whole growth stages of oilseed can be predicted by this method with an r² of 0.983, RMSE of 2.60 cm, and rRMSE of 7.14%. Furthermore, this method enabled a comprehensive Genome-wide association study (GWAS) in a 293-accession genetic population. The GWAS identified 200 and 65 statistically significant single nucleotide polymorphisms (SNPs), which were tightly associated with 28 and 11 candidate genes, at the late bolting and flowering stages, respectively. These findings demonstrated that the proposed method is promising for accurate estimations of IHs in oilseed rape as well as exploring the variations within the subplot, thus providing great potential for high-throughput plant phenotyping in crop breeding.

Shadows are inevitable in vegetated remote sensing scenes due to variations in viewing and solar geometries, resulting in illuminated vegetation, shadowed vegetation, illuminated background and shadowed background. In RGB images, shadowed vegetation is difficult to separate from the shadowed background because their spectra are very similar in the visible light range. Furthermore, shadowed vegetation may provide different ecological functions than illuminated vegetation. Therefore, it is important to extract both illuminated and shadowed vegetation instead of combining them into one vegetation class. However, most previous studies focused on extracting total vegetation cover and neglected separating illuminated and shadowed vegetation, partly due to a lack of sufficient information. In this study, polarization information is introduced to extract illuminated vegetation, shadowed vegetation and background simultaneously with different deep learning algorithms. The experimental results show that the addition of polarization information can effectively improve the extraction accuracy of illuminated vegetation, shadowed vegetation and background, with a maximum accuracy improvement of 12.2%. The accuracy of shadow vegetation improved the most, with a rate of 21.8%. The results of this study suggest that by adding polarization information, illuminated and shadowed vegetation can be accurately extracted to provide a reliable vegetation cover product for remote sensing.