Vineyard yield prediction method and system based on Internet of Things

Abstract

The invention discloses a vineyard yield prediction method and system based on the Internet of Things, and relates to the technical field of intelligent agriculture. The method comprises the following steps: deploying intelligent sensing equipment in a park to collect air temperature and humidity, illumination, wind speed, soil moisture and salinity data in real time, and synchronously utilizing an RTK-GPS unmanned aerial vehicle to obtain multispectral and RGB remote sensing images; and performing pixel-level alignment and fusion on the growth microenvironment data and the remote sensing image in a space-time dimension to generate a multi-channel input tensor containing a sensor numerical channel and an image channel, constructing a CNN model by taking an expanded convolutional network or a U-Net variant as a core, directly outputting a high-resolution pixel-level cluster density heat map, and obtaining a high-resolution cluster density heat map. The high-precision yield prediction result is obtained through global single spike weight coefficient conversion, the yield prediction error caused by a traditional manual sampling method can be reduced, meanwhile, the labor cost is reduced, and the fine management level of a vineyard is improved.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method and system for predicting vineyard yield based on the Internet of Things. Background Technology

[0002] Vineyard yield forecasting is a core component of precision agriculture and smart grape cultivation, as its accuracy directly determines harvesting plans, resource allocation, market pricing, and supply chain efficiency. Traditional forecasting methods primarily rely on manual field sampling, extrapolating the overall vineyard yield by manually counting and measuring bunch weights. This method is labor-intensive, time-consuming, and significantly influenced by subjective experience, with errors often ranging from 10% to 30%, making it difficult to meet the demands of modern vineyards for refined management that emphasizes early maturity, high yield, and high quality.

[0003] With the rise of Internet of Things (IoT) technology, low-cost sensors, low-power wide-area communication (LoRaWAN, NB-IoT), and edge computing gateways are gradually being implemented in agricultural scenarios. Existing technologies have attempted to upload environmental data such as soil moisture, air temperature and humidity, and light intensity to the cloud, and then estimate yield through statistical regression or simple machine learning models. However, these solutions generally suffer from the following shortcomings: 1) Single data source: Relying solely on ground-based point sensors, lacking coupling with canopy spatial distribution information, resulting in low spatial resolution and difficulty in capturing differences between plants. 2) Feature fragmentation: Remote sensing images and sensor data are usually processed independently, without deep integration at the "pixel-environment" level, preventing the model from simultaneously utilizing spectral reflectance features and micro-environmental variables. 3) Poor model adaptability: Traditional regression models have limited ability to characterize nonlinear, high-dimensional interactions; prediction accuracy rapidly declines when weather, soil, variety, or management practices change. 4) Lack of real-time performance and closed-loop processing: Most systems only perform offline analysis and cannot dynamically update with new data, leading to long-term accumulation of "model drift."

[0004] In recent years, Convolutional Neural Networks (CNNs) have made groundbreaking progress in the field of computer vision, and their powerful spatial feature extraction capabilities have been tested for tasks such as fruit tree detection and lesion identification. However, most existing research remains at the level of "visible light / multispectral imagery → target detection / segmentation," and has not yet formed a complete technical closed loop of "ground IoT environmental data + UAV remote sensing imagery → pixel-level yield density map → whole-orchard yield prediction." In particular, existing literature has not provided systematic and engineeringable solutions for how to perform pixel-level alignment of heterogeneous spatiotemporal data, how to construct high-dimensional input tensors, how to explicitly introduce sensor variables into the CNN architecture, and how to establish an edge-cloud collaborative framework that can evolve online.

[0005] In summary, current vineyard yield prediction technologies suffer from bottlenecks such as fragmented data, coarse models, poor real-time performance, and insufficient spatial granularity. Therefore, there is an urgent need for a method and system that can integrate multi-source sensor data from the Internet of Things (IoT) with high-resolution UAV imagery, and achieve high accuracy, high spatial resolution, and online updates through deep learning models. To this end, an IoT-based vineyard yield prediction method and system is proposed. Summary of the Invention

[0006] The main objective of this invention is to provide a method and system for predicting vineyard yield based on the Internet of Things (IoT). By integrating multi-source sensor data from the IoT with high-resolution imagery from drones, and using a deep learning model, this invention achieves high-precision, high-spatial-resolution, and online-updable vineyard yield prediction results, effectively solving the problems in the background technology.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting vineyard yield based on the Internet of Things (IoT) includes: The vineyard is divided into several management units according to soil type, variety and topography. Each unit is assigned a unique ID and an RTK-GPS reference point is established. Growth microenvironment data of the center location of all the management units are collected in real time. Using drones to acquire multispectral and RGB remote sensing images of vineyards; The growth microenvironment data and the remote sensing image are aligned and fused at the pixel level in the spatiotemporal dimension to generate a multi-channel input tensor containing sensor numerical channels and image channels. Using the multi-channel input tensor as input and the pixel-level ear density heatmap as output, a convolutional neural network model is constructed and trained. The pixel-level cluster density heatmap is converted into vineyard yield prediction results.

[0008] A vineyard yield prediction system based on the Internet of Things includes: A growth microenvironment data acquisition module is installed at the center of the vineyard's management unit to collect real-time growth microenvironment data at the center of the management unit. The UAV remote sensing module is used to acquire multispectral and RGB remote sensing images of the vineyard. An edge computing gateway is used to receive the growth microenvironment data collected by the growth microenvironment data acquisition module and the remote sensing image collected by the UAV remote sensing module in real time, and to perform pixel-level alignment and fusion of the growth microenvironment data and the remote sensing image in the spatiotemporal dimension to generate a multi-channel input tensor containing sensor numerical channels and image channels. User-side intelligent devices are used to acquire and upload the actual output data of the management unit through manual sampling. A cloud management platform is used to receive the multi-channel input tensor and the actual yield data, and to construct and train a convolutional neural network model with the multi-channel input tensor as input and the pixel-level ear density heatmap as output, so as to convert the pixel-level ear density heatmap into vineyard yield prediction results.

[0009] Furthermore, the growth microenvironment data includes one or more of the following: air temperature and humidity, light intensity, wind speed, soil moisture content, and soil salinity.

[0010] Furthermore, the specific steps for pixel-level alignment and fusion are as follows: Using the time of drone capture as the time reference, linear interpolation or inverse distance weighted interpolation is performed on the sensor data to generate a sensor numerical map with the same spatial resolution as the remote sensing image. The sensor numerical image is stitched together with RGB, NDVI images and position-coded images to form a multi-channel input tensor that includes sensor numerical channels and image channels.

[0011] Furthermore, the multi-channel input tensor is composed of the following channels concatenated in sequence: RGB channel, NDVI channel, sensor channel, and two-dimensional position encoding channel.

[0012] Furthermore, the convolutional neural network model is an dilated convolutional network or a variant of U-Net, wherein the dilation rate sequence of the dilated convolutional network is [1,2,4,8], and the U-Net variant uses dilated convolution instead of standard convolution in the bottleneck layer.

[0013] Furthermore, the output of the convolutional neural network model is reduced to a single channel using 1×1 convolution, and then a 0–1 normalized ear density heatmap is generated by passing it through the Sigmoid activation function. Finally, it is multiplied by the global single ear weight coefficient to obtain the absolute yield map.

[0014] Furthermore, the training process of the convolutional neural network model uses the product of the actual number of manually sampled ears of fruit and the average weight of the ears of fruit as the true label, and the loss function is the weighted mean square error, wherein the weight of the ear of fruit region is 5–10 times that of the background region. Furthermore, the training optimizer used in the training process is AdamW, with an initial learning rate of 1e-3, and cosine annealing scheduling. During training, random rotation, flipping, brightness and contrast perturbations are used to perform online data augmentation on the 11-channel input tensor.

[0015] Furthermore, the training process further includes a hyperparameter optimization step, wherein the hyperparameter optimization step uses a Bayesian optimization algorithm to search within the hyperparameter space composed of the expansion rate combination, the number of residual blocks, the channel cardinality, and the weighted loss background weights, and the objective function is the validation set mean square error.

[0016] The present invention has the following beneficial effects: Compared with existing technologies, this solution achieves pixel-level deep fusion of IoT environment-soil data and high-resolution drone imagery. The CNN model can simultaneously learn canopy spectral characteristics and microenvironmental differences, greatly reducing yield prediction errors and workload caused by traditional manual sampling methods, and meeting the needs of commercial precision harvesting. Compared with existing technologies, this solution can directly locate high-yield / low-yield areas by outputting pixel-level cluster density heat maps, which is beneficial for guiding variable irrigation, fertilization and manual fruit thinning, and realizes "one policy per plant" refined management to improve the overall quality consistency. Compared with existing technologies, this solution utilizes Internet of Things (IoT) technology and employs intelligent sensing devices to collect data, enabling the detection of potential yield reduction risks during critical windows such as flowering and color-changing periods. This facilitates the timely initiation of frost protection, supplemental lighting, or deficit irrigation measures.

[0017] Compared with existing technologies, this solution decouples the sensor nodes from the CNN architecture. When a new sensor is needed, such as CO2 or stem water potential, only channels need to be added without reconstructing the network, which has good versatility and scalability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process of a vineyard yield prediction method based on the Internet of Things according to the present invention; Figure 2 This is a schematic diagram of the structure of a vineyard yield prediction system based on the Internet of Things according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] A method for predicting vineyard yield based on the Internet of Things (IoT) includes: Real-time acquisition of growth microenvironment data at the center of all management units; Using drones to acquire multispectral and RGB remote sensing images of vineyards; The growth microenvironment data and remote sensing images are aligned and fused at the pixel level in the spatiotemporal dimension to generate a multi-channel input tensor that includes sensor numerical channels and image channels. A convolutional neural network model was constructed and trained using a multi-channel input tensor as input and a pixel-level ear density heatmap as output. Convert pixel-level cluster density heatmaps into vineyard yield prediction results.

[0021] In one possible embodiment, the specific process of the technical solution of the present invention includes the following steps: The vineyards are divided into several management units based on soil type, variety, and topography. Each unit is assigned a unique ID and an RTK-GPS reference point is established.

[0022] LOT Node Deployment and Networking A growth microenvironment data acquisition module is deployed at the center of each management unit to collect data on air temperature and humidity, light intensity, wind speed, soil moisture, and salinity; the frequency of data upload by the nodes is set to once every 15 minutes.

[0023] UAV remote sensing image acquisition With a preset flight path of 80% overlap, orthophotos were captured using a multispectral + RGB camera during the grape's flowering period, veraison period, and three weeks before harvest, and RTK coordinates were recorded simultaneously.

[0024] Data preprocessing and synchronization The edge computing gateway performs anomaly removal, interpolation, and normalization on IoT time-series data. Radiometric correction, geometric correction, and noise reduction are performed on the drone images, and the images are cropped into fixed-size tiles according to the management unit. In this embodiment, 512×512px is used as an example. By using LOT variables within the same spatiotemporal window as channel-level features of the image, "image + sensor" sample pairing is achieved.

[0025] CNN Model Building and Training Choose either the dilated convolutional network or the U-Net variant, with a multi-channel image as input and a pixel-level heatmap of ear density as output; Using manually sampled actual output as labels, mean squared error (MSE) as the loss function, and Adam optimizer for training, with 80% of the data used for training and 20% for validation, hyperparameters were searched using Bayesian optimization.

[0026] After training is completed, real-time ear counting will be achieved; The cloud management platform retains the full-precision model for weekly rolling retraining to incorporate the latest observation data.

[0027] Production forecasting and spatial mapping Multiply the density heatmap output by the CNN by the average single ear weight to obtain the yield per plant / row; Kriging interpolation was used to generate a full-plant yield distribution map, and soil and topography layers were overlaid on a GIS platform to form a visual decision map.

[0028] Model closure and iteration The forecast results are then pushed to irrigation, fertilization, and labor dispatch systems; After harvesting, the actual yield is compared with the model to calculate RMSE and MAE, triggering incremental learning or model retraining to achieve continuous accuracy improvement.

[0029] Through the above steps, vineyards can obtain high-precision yield predictions throughout the growing season, while using LOT-CNN closed loop to reduce manual sampling costs and improve resource utilization efficiency.

[0030] The methods and approaches for acquiring multi-channel images are as follows: RGB remote sensing images The visible light camera (RGB sensor) carried by the drone directly captures raw 3-band images during the shooting route; Subsequently, orthorectification, radiometric calibration, and stitching are performed to generate 8-bit or 12-bit three-channel orthophotos, namely Red, Green, and Blue.

[0031] NDVI images The raw multispectral images were acquired using the multispectral sensor carried by the drone, which includes both near-infrared (NIR) and red bands. After radiometric calibration, orthorectification and stitching of the near-infrared (NIR) and red bands, the NDVI is calculated using the formula: NDVI = (NIR - Red) / (NIR + Red) to obtain a single-channel NDVI floating-point image, which is then linearly mapped to 8-bit for storage.

[0032] Location coding map During the georegistration process of generating orthophotos, the X and Y coordinates of each pixel in the target projection coordinate system are calculated using the RTK-GPS exterior orientation elements of each image. After normalizing X and Y to the [-1,1] interval, a two-dimensional floating-point position encoding map (X-map, Y-map) with the same width and height as the image is generated.

[0033] The growth microenvironment data and UAV images are fused into a single "multichannel tensor" and then used as input to the CNN model. The specific steps include: alignment, encoding, fusion, and training. 1) Spatiotemporal alignment 1.1) Time alignment: Based on the UAV takeoff time, the latest observation value of the LOT node within the past 6 hours is taken as the "instantaneous field"; if the node sampling interval is >6 hours, it is linearly interpolated to the takeoff time.

[0034] 1.2) Spatial Alignment For each 512×512 pixel image tile, obtain the GPS coordinates of its center point; Using inverse distance weighting, all node readings within a radius R are interpolated into five 512×512 "sensor images", with a resolution completely consistent with RGB / NDVI.

[0035] Result: RGB(3) + NDVI(1) + 5 sensor images = 9-channel input, size 512×512×9.

[0036] 2) Channel encoding 2.1) Normalization: RGB→[0,1]; NDVI→[-1,1]; Subtract the mean from each sensor channel and divide by the standard deviation.

[0037] 2.2) Location coding: Two additional coordinate graphs (x, y normalized to [-1, 1]) are generated as the 10th and 11th channels to help the network learn spatial bias.

[0038] The final input is 512×512×11.

[0039] 3) Internal network integration 3.1) Early fusion: Simply feed all 11 channels into the first convolutional layer.

[0040] 3.2) Intermediate Fusion (Optional): If there are concerns about sensor noise, the RGB / NDVI can be routed through a CNN branch, and the sensor image can be upgraded to 1×1 Conv before concatenation. Experiments have shown that the difference is <1% MAE.

[0041] 3.3) Attention weighting: Add an SE-block at the bottleneck of U-Net to allow the network to automatically learn the importance of the channel and avoid manually setting weights.

[0042] 4) Training Process 4.1) Sample: The density heatmap of each ear of fruit obtained by manual sampling is used as a 1-channel label (512×512), which corresponds one-to-one with the input tensor.

[0043] 4.2) Loss: MSE (ear density) + 0.1×MAE (global yield) are jointly monitored to ensure simultaneous optimization at the pixel level and the whole plant level.

[0044] 4.3) Data augmentation: Flip / rotate / perturb the 11 channels as a whole, and synchronously perform affine transformation on the sensor numerical channels to maintain spatial consistency.

[0045] 4.4) Training: AdamW, lr=1e-3, cosine annealing, batch=8, 150 epochs, early stop patience=15.

[0046] 4.5) Inference: If the same 11-channel tensor is forwarded once, the network outputs a 512×512×1 density heatmap, which is multiplied by the single ear weight coefficient to obtain the yield map.

[0047] In one possible implementation, the CNN model building and training process can be broken down into the following steps: Channel design The drone orthophoto was sliced ​​into 512×512px tiles, while the RGB three channels remained unchanged; The NDVI is calculated to obtain the 4th channel; The latest 6-hour average values ​​(temperature, humidity, soil moisture, light intensity, and wind speed) of 5 LOT nodes within a 3m×3m range at the center of the same tile are transformed into 5 "sensor images" of 512×512 using bilinear interpolation, ultimately resulting in 9-channel input.

[0048] Two additional coordinate graphs (x, y normalized to [-1, 1]) are generated as inputs for channels 10 and 11.

[0049] Final input: RGB(3)+NDVI(1)+LOT(5)+L(2); i.e. 512×512×11.

[0050] Label making During manual sampling, the center coordinates of each plant were recorded using RTK-GPS; If the number of ears n and the average weight w of 10 ears are measured on-site, then the yield of the plant y = n × w; The y-value is rasterized at a resolution of 0.25m to form a density heatmap of the same size as the tile, with a background value of 0 and a Gaussian core (σ=0.3m) diffusion at the center of the ear to ensure that the label is continuous and guideable.

[0051] Dataset partitioning Random stratified sampling is performed by management unit: 80% of the tiles are used for training and 20% for validation; at the same time, an entire plot of land that did not participate in the training is reserved as the final blind test area to prevent spatial leakage.

[0052] Data Augmentation During the training phase, random horizontal flipping, 90° rotation, ±10% brightness change, ±5% contrast change, and random Gaussian noise σ∈[0,2] are performed online. The LOT channel is synchronously subjected to the same geometric transformation, and the numerical channel is subjected to ±5% multiplicative noise to simulate sensor drift.

[0053] Network selection If a greater focus is placed on the global receptive field, an expanded convolutional network can be chosen: a backbone of 4 residual blocks, 3×3 convolutional kernels, expansion ratio [1,2,4,8], and 64 output channels per layer; For fine segmentation, select the U-Net variant: the encoder downsamples 4 times (64→128→256→512), the decoder upsamples 4 times and adds skip connections, the bottleneck layer dilated convolution replaces the ordinary convolution, and the output is maintained at 512×512.

[0054] Output head design Add a 1×1 convolution at the end of the network to reduce the dimensionality to 1 channel, then add a Sigmoid function to restrict the range of 0–1 to represent the relative density; multiply by the global coefficient α (average single ear weight of the training set) to obtain the absolute yield g / px heatmap.

[0055] loss function Using MSE: To suppress background dominance, additional weighting is applied: pixels with label values ​​greater than 0 are multiplied by 10, and the background is multiplied by 1, so that the network focuses on learning the ear region.

[0056] Optimizer and Learning Rate Adam, Initial The weight decays by 1e-4; Training for 150 epochs, cosine annealing to 1e-5, and linear warm-up for the first 5 epochs.

[0057] Bayesian hyperparameter search Search space: Expansion rate combination: [(1,2,4,8),(1,2,5,9),(1,3,6,9)] Number of residual blocks: [3,4,5] Channel base: [32, 64, 96] Weighted MSE background weights: [1,2,5,10] Using the Optuna library, the goal is to validate MSE by running 50 trials with an early stop patience of 10 epochs.

[0058] Training monitoring Record train / valMSE, MAE, and R² for each epoch; Real-time visualization of the difference between predicted heatmaps and ground truth values ​​using TensorBoard; If valMSE does not decrease for 10 consecutive epochs, the optimal weight will be automatically rolled back.

[0059] Model compression and deployment After selecting the optimal model, quantization to INT8 using TensorRT resulted in a mAP decrease of <1%. In NVIDIA Jetson Xavier gateway inference, a single 512×5129 channel image takes less than 60ms, enabling real-time tracking of drone video streams.

[0060] Online closed-loop update After each harvest, new sampling points are added to the database, triggering incremental training: the encoder weights are fixed, and only the last two layers are fine-tuned for 5 epochs with a learning rate of 1e-4 to prevent catastrophic forgetting. A full retraining is performed quarterly to maintain the model's timeliness.

[0061] The above steps are used to predict vineyard yields using dilated convolutional networks or U-Net variants.

[0062] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting vineyard yield based on the Internet of Things, characterized in that, include: The vineyard is divided into several management units according to soil type, variety and topography. Each unit is assigned a unique ID and an RTK-GPS reference point is established. Growth microenvironment data of the center location of all the management units are collected in real time. Using drones to acquire multispectral and RGB remote sensing images of vineyards; The growth microenvironment data and the remote sensing image are aligned and fused at the pixel level in the spatiotemporal dimension to generate a multi-channel input tensor containing sensor numerical channels and image channels. Using the multi-channel input tensor as input and the pixel-level ear density heatmap as output, a convolutional neural network model is constructed and trained. The pixel-level cluster density heatmap is converted into vineyard yield prediction results.

2. The vineyard yield prediction method based on the Internet of Things according to claim 1, characterized in that, The growth microenvironment data includes one or more of the following: air temperature and humidity, light intensity, wind speed, soil moisture content, and soil salinity.

3. The vineyard yield prediction method based on the Internet of Things according to claim 1, characterized in that, The specific steps for pixel-level alignment and fusion are as follows: Using the time of drone capture as the time reference, linear interpolation or inverse distance weighted interpolation is performed on the sensor data to generate a sensor numerical map with the same spatial resolution as the remote sensing image. The sensor numerical image is stitched together with RGB, NDVI images and position-coded images to form a multi-channel input tensor that includes sensor numerical channels and image channels.

4. The vineyard yield prediction method based on the Internet of Things according to claim 3, characterized in that, The multi-channel input tensor is composed of the following channels concatenated in sequence: RGB channel, NDVI channel, sensor channel, and two-dimensional position encoding channel.

5. A method for predicting vineyard yield based on the Internet of Things according to any one of claims 1-4, characterized in that, The convolutional neural network model is a dilated convolutional network or a variant of U-Net. The dilated convolutional network has a dilation rate sequence of [1,2,4,8]. The U-Net variant uses dilated convolution instead of standard convolution in the bottleneck layer.

6. The vineyard yield prediction method based on the Internet of Things according to claim 5, characterized in that, The output of the convolutional neural network model is reduced to a single channel using 1×1 convolution, and then a 0–1 normalized ear density heatmap is generated by the Sigmoid activation function. Finally, it is multiplied by the global single ear weight coefficient to obtain the absolute yield map.

7. The vineyard yield prediction method based on the Internet of Things according to claim 6, characterized in that, The training process of the convolutional neural network model uses the product of the actual number of manually sampled ears of fruit and the average weight of the ears as the true label, and the loss function is the weighted mean square error, where the weight of the ear region is 5–10 times that of the background region.

8. The vineyard yield prediction method based on the Internet of Things according to claim 6, characterized in that, The training optimizer used in the training process is AdamW, with an initial learning rate of 1e-3. Cosine annealing scheduling is employed, and random rotation, flipping, brightness and contrast perturbations are used to perform online data augmentation on the 11-channel input tensor during training.

9. A vineyard yield prediction method based on the Internet of Things according to claim 8, characterized in that, The training process further includes a hyperparameter optimization step, wherein the hyperparameter optimization step uses a Bayesian optimization algorithm to search within the hyperparameter space composed of the expansion rate combination, the number of residual blocks, the channel cardinality, and the weighted loss background weights, and the objective function is the validation set mean square error.

10. A vineyard yield prediction system based on the Internet of Things, characterized in that, include: A growth microenvironment data acquisition module is installed at the center of the vineyard management unit to collect growth microenvironment data in real time, including one or more combinations of air temperature and humidity, light intensity, wind speed, soil moisture content and soil salinity. The UAV remote sensing module is used to acquire multispectral and RGB remote sensing images of the vineyard. An edge computing gateway is used to receive the growth microenvironment data collected by the growth microenvironment data acquisition module and the remote sensing image collected by the UAV remote sensing module in real time, and to perform pixel-level alignment and fusion of the growth microenvironment data and the remote sensing image in the spatiotemporal dimension to generate a multi-channel input tensor containing sensor numerical channels and image channels. User-side intelligent devices are used to acquire and upload the actual output data of the management unit through manual sampling. A cloud management platform is used to receive the multi-channel input tensor and the actual yield data, and to construct and train a convolutional neural network model with the multi-channel input tensor as input and the pixel-level ear density heatmap as output, so as to convert the pixel-level ear density heatmap into vineyard yield prediction results.

Images