[Paper Summary] Essay notes on the application of deep learning in the field of agriculture

Table of Contents of Articles

• 1. Deep learning

• 1.1 Image Classification

• 1. A Neural Network Method for Classification of Sunlit and Shaded Components of Wheat Canopies in the Field Using High-Resolution Hyperspectral Imagery
• 2. Crop pest classification with a genetic algorithm-based weighted ensemble of deep convolutional neural networks
• 3. Data augmentation for automated pest classification in Mango farms
• 4.Data augmentation for automated pest classification in Mango farms
• 5. An attribution-based pruning method for real-time mango detection with YOLO network
• 6. Predicting pasture biomass using a statistical model and machine learning algorithm implemented with remotely sensed imagery
• 7.Irrigation water infiltration modeling using machine learning
• 8. Short term soil moisture forecasts for potato crop farming: A machine learning approach
• 9. Wild blueberry yield prediction using a combination of computer simulation and machine learning algorithms
• 10. Citrus advisory system: A web-based postbloom fruit drop disease alert system
• 11. Automated crop plant counting from very high‑resolution aerial imagery
• 12. EfficientNet-B4-Ranger: A novel method for greenhouse cucumber disease recognition under natural complex environment
• 13.Automatic vegetable disease identification approach using individual lesion features
• 14. Comparison of convolution neural networks for smartphone image based real time classification of citrus leaf disease

• 1.2 Semantic segmentation

• 1. Establishing a model to predict the single boll weight of cotton in northern Xinjiang by using high resolution UAV remote sensing data
• 2. Tomato leaf segmentation algorithms for mobile phone applications using deep learning
• 3. Biophysical parameters of coffee crop estimated by UAV RGB images
• 4. Vineyard yield estimation by combining remote sensing, computer vision and artificial neural network techniques
• 5. Accuracy of carrot yield forecasting using proximal hyperspectral and satellite multispectral data
• 6. Evaluation of cotton emergence using UAV-based imagery and deep learning
• 7. Deep learning techniques for automatic butterfly segmentation in ecological images
• 8. Improving segmentation accuracy for ears of winter wheat at flowering stage by semantic segmentation Semantic segmentation was used to improve the segmentation accuracy of winter wheat ears at flowering stage

• 2.3 Object Detection

• 1. Brown rice planthopper (Nilaparvata lugens Stal) detection based on deep learning
• 2. Identification of olive fruit, in intensive olive orchards, by means of its morphological structure using convolutional neural networks

• 2. Machine learning

• 1. A random forest ranking approach to predict yield in maize with uav-based vegetation spectral indices

• 3. Miscellaneous

• 1. DropLeaf: A precision farming smartphone tool for real-time quantification of pesticide application coverage
• 2. Underutilised crops database for supporting agricultural diversification
• 3. Replacing traditional light measurement with LiDAR based methods in orchards
• 4. A cyber-physical intelligent agent for irrigation scheduling in horticultural crops
• 5. Biophysical parameters of coffee crop estimated by UAV RGB images
• 6. A satellite-based ex post analysis of water management in a blueberry orchard
• 7.A 3D white referencing method for soybean leaves based on fusion of hyperspectral images and 3D point clouds
• 8. Assessing winter wheat foliage disease severity using aerial imageryacquired from small Unmanned Aerial Vehicle (UAV)

1. Deep learning

1.1 Image Classification

1. A Neural Network Method for Classification of Sunlit and Shaded Components of Wheat Canopies in the Field Using High-Resolution Hyperspectral Imagery

Summary: This paper uses the method of image classification. Linear discriminant analysis (LDA) to find linear combinations of features that characterize the pre-selected categories (Shadow Leaf (SHL), Shadow Ear (SHE), Sunlight Leaf (SL), Sunlight Ear (SE), and Background (BG). The RGB data consists of three bands: 620 nm (red), 535 nm (green), and 445 nm (blue). The annotated data were collected from hyperspectral images of 23 wheat varieties at different crop growth stages. The total number of annotation data for the 5 categories of SL, SE, SHL, SHE, and BG is 119,447, 164,223, 11,644, 4361, and 227,232 pixels, respectively (this place should refer to photos). The CNN model was compared to gradient descent and support vector machine classification using mean accuracy (AA), F-score, and recall score. Advantages: The study consisted of an actual filming device, with adequate field experiments and a large data set. Disadvantages: After reading this paper, I don't know what kind of practical problem is solved. Reference value: 2 points.

2. Crop pest classification with a genetic algorithm-based weighted ensemble of deep convolutional neural networks

Summary: Throughout the study, seven different pre-trained CNN models (VGG-16, VGG-19, ResNet-50, Beginning-v3, Xception, MobileNet, SqueezeNet) modified and retrained the public D0 dataset for 40 classes using appropriate transfer learning and fine-tuning strategies. Subsequently, the three best-performing CNN models, Inception-V3, Xception, and MobileNet, were integrated with maximum probability and strategy to improve classification performance, and the model was named SMPEnsemble. These models are then integrated using a weighted voting method.

The genetic algorithm comprehensively considers the success rate and prediction stability of the three CNN models, determines the weight, and the model is named GAEnsemble. GAEnsemble achieved the highest classification accuracy of 98.81% on the D0 dataset. In order to enhance the robustness of the ensemble model, the process was repeated with two other datasets, namely the small dataset of class 10 and the IP102 dataset of class 102, without changing the CNN model with the best initial performance on D0. The accuracy of GAEnsemble is 95.15% for small datasets and 67.13% for IP102. Pros: Not outstanding, the paper is very ordinary. Disadvantages: The dataset is small. Reference value: 2 points.

3. Data augmentation for automated pest classification in Mango farms

Summary: This paper proposes an advanced machine learning (ML) technique for analyzing large-scale mango fields and using computer vision and deep learning techniques to identify the occurrence of biological threats. The ML technique proposed in this paper extends the pre-trained VGG-16 deep learning model to supplement the last layer with a two-layer fully connected network training. In addition, this study also takes into account the practical operating conditions faced by Indonesian farmers when collecting and processing visual information from mango farms. The sparsity of dataset availability for effectively training deep learning networks is addressed by applying a data augmentation process that accurately reproduces the conditions faced by farmers. On the validation dataset and the test dataset, the overall accuracy of the training scheme proposed in this paper is 73% and 76%, respectively. After applying the augmentation transform function, the accuracy of the test data is improved by 13.43%.

4.Data augmentation for automated pest classification in Mango farms

Summary: Mango pest classification framework, which consists of 15 categories, according to the improvement of the VGG-16 network, divides 16 types of pests and healthy leaves. The proposed ML technique extends the pre-trained VGG-16 deep learning model to supplement the last layer with a two-layer fully connected network training. The sparsity of the availability of datasets for effectively training deep learning networks is addressed by applying a data augmentation process that accurately reproduces the conditions faced by farmers and receives classified outputs in real-time for pest categories that may affect mango production. Pros: Finally, an Android-based APP was developed. Cons: I don't know. Personally, I feel that there are no bright spots, and there is only one web interface, and it is not said that the developed APP solves any problems. Reference value: 3 points.

5. An attribution-based pruning method for real-time mango detection with YOLO network

Summary: This study proposes a generalized attribution method for pruning detection network that is easy to fine-tune to detect mango. By designing channels and spatial masks to generalize attribution methods, convolutional kernels in the original YOLOv3-tiny network that are closely related to the output of a specific target can be detected. Then, uncorrelated nuclei are pruned layer by layer for channel dimension. Before fine-tuning the pruned network, anchor size, data augmentation, and learning rate attenuation were used to detect mangoes. Experimental results show that the obtained network is a mango detection network with invariant scale and rotation, and obtains an f1 score of 0.944 under 2.6 GFLOPs (giga-floating-point operation). Compared to fine-tuned networks without pruning, our network is 68.7% less computational and 0.4% more accurate.

Compared with the state-of-the-art network trained with the same Mango dataset, the computational effort of the algorithm is reduced by 83.4%, and the accuracy loss is only about 2.4%. The proposed pruning method can strip a subnet from the large-scale detection network to meet the real-time requirements of low-power processors of mobile devices, such as the ARM Cortex-A8 to perform about 4.0 GFLOPS (giga-floating-point operations per second). Trained networks and test code are available for comparative study**.

Pros**: This paper was submitted in July 2019. Fine-tuning and pruning are proposed. Cons: Oddly said it was a mango test, but I didn't find the mango dataset in the paper, and there was no field experiment. It's more like a general engineering paper, and I want to send it to a journal with good agronomy. Reference value: 2 points, similar to what we did, but not as good as what we did.

6. Predicting pasture biomass using a statistical model and machine learning algorithm implemented with remotely sensed imagery

Summary: The aim of this study was to test the performance of an ensemble method combined with the obtained remote sensing images with multispectral cameras mounted on unmanned aerial vehicles (UAVs), statistical models (GAM total additive models) and machine learning algorithms (random forests, RF) implemented with publicly available data to predict future forage biomass loads. This study showed that both GAM and RF models could predict grassland biomass yield before grazing by using grassland growth observations, environmental variables and grassland management variables, with an average error of less than 20%. Pros: The paper has actual weather data, fertilization data, as well as drone flight data. Disadvantages: Only the training set and the test set. No new methods were used. Reference value: 2 points.

7.Irrigation water infiltration modeling using machine learning

Summary: In this study, five standard AI models are proposed, including Artificial Neural Network (ANN), Adaptive Neural Fuzzy Reasoning System (ANFIS), Grouped Data Processing Method (GMDH), Multiple Linear Regression, and Support Vector Regression (SVR), as well as their comprehensive models combined with the Firefly Algorithm (FA) to predict the infiltration water in the ditch of the irrigation system. FA is an optimization tool when building a comprehensive model. Data to evaluate these models were collected from published literature and field experiments conducted at research farms at Komansbuk University in Iran. The input parameters of the model were ditch length (L), infiltration rate (Q), ditch tail advance time (TL), infiltration cross-sectional area (Ao) and infiltration opportunity time (To). Root mean square (RMSE), mean absolute error (MAE), correlation coefficient (R2), Nash-Sutcliffe efficiency index (NSE) and consistency index (IA) were used to evaluate the prediction effect of the model.

The results show that FA can improve the accuracy of the model, and the RMSE values in ANFIS, GMDH, MLPNN and SVR are increased by 5%, 1%, 4% and 47%, respectively. The calculated composite index (SI) shows that the combination of SVR and FA significantly improves the performance of the standard SVR model by 97%. Advantages: Irrigation systems that use a variety of models and FA algorithms to predict water. Cons: Dataset size is not shown. Reference value: 3 points.

8. Short term soil moisture forecasts for potato crop farming: A machine learning approach

Summary: This paper describes experiments using sensors using 3 years of data in a variety of scenarios. 3 soil depths, measured 10 times a year. and weather change data as well as data from other literature. Advantages: Abundant experimental data. Disadvantages: Compared to big data, there is still less data. The methods used are feature extraction, support vector machines, and neural networks. Reference value: 3 points.

9. Wild blueberry yield prediction using a combination of computer simulation and machine learning algorithms

Summary: In this study, the wild blueberry pollination model generated by the data we used is a spatially explicit simulation model, which is validated by field observations and experimental data in Maine, USA for nearly 30 years (simulation model). The main objective of this study was to evaluate the relative importance of bee species composition and weather factors in regulating wild blueberry agroecosystems. Specifically, we sought to uncover how bee species composition and weather affect yields, and predict optimal yields. Honey bee species composition and weather conditions using computer simulations and machine learning algorithms to achieve optimal yields.

Multiple Linear Regression (MLR), Enhanced Decision Tree (BDT), Random Forest (RF), and Extreme Gradient Enhancement (XGBoost) were evaluated as prediction tools. We also made a predictor selection before submitting data to the learning algorithm. In this way, we can reduce the number of dimensions of the input without significantly reducing the accuracy of the prediction. The results showed that clone size, honeybee, bumblebee, male bee species, female species, maximum high temperature range, and number of precipitation days were the optimal variables for prediction. The results showed that XGBoost outperformed other algorithms in the prediction performance of all measures and the yield of wild blueberries reached a certainty coefficient (R2) of 0.938, a root mean square error (RMSE) of 343.026, and a mean absolute error (MAE) and a relative root mean square error of 5.444% in 206. Pros: Simulation model used. Disadvantages: I don't feel innovative in the method. Reference value: 4 points.

10. Citrus advisory system: A web-based postbloom fruit drop disease alert system

Summary: PFD (Postharvest Orange Disease) may cause fruit drop after flowering in citrus is a serious fungal disease of citrus that can cause premature fruit shedding. The study developed a web-based tool to assist citrus growers in their spraying decisions for managing PFD risk. Information technologies such as databases, queries, and programming languages have been used to develop this tool. The system collects weather data from weather stations installed by the Florida Automatic Weather Network (FAWN) and the Agroclimatology Research Group and uses weather observations to run PFD disease models and estimate how favorable the environment is for infection. The system sends notifications to farmers and recommends the use of fungicides based on different PFD risks and flowering periods. First, the weather data source, leaf humidity model and leaf humidity decision algorithm are described.

Finally, we describe how information technology can be used to provide solutions that allow users to easily access the system. Pros: Web interface. The system is currently in use in Florida, but the authors aim to expand its geography to other citrus-growing states in the United States. Disadvantages: The method is very common, which is equivalent to using the climate data area to simulate the leaf humidity model, and then calculate the time interval index in the PFD model, when the conidia germination index reaches a certain threshold, it will promote the spray to achieve the purpose of preventing PDF. Reference value: 4 points. What can we predict if we have weather data?

11. Automated crop plant counting from very high‑resolution aerial imagery

Summary: In this study, an automated method for counting plants from ultra-high-resolution drone images is proposed. The proposed method uses machine vision – the excess green index and Otsu's method – and uses convolutional neural networks for transfer learning to identify and count plants. An ensemble method has been implemented to count 10-week-old spinach plants in an experimental field with a surface area of 3.2 hectares. Validation data for plant counts can be used for 1/8 of the surface area. The results show that the proposed methodology can count plants with a spatial resolution of 8 mm/pixel in an area of up to 172 m2 with an accuracy of 95%. In addition, when the spatial resolution is reduced by 50%, the maximum additional counting error obtained is 0.7%. In the end, a total error of 42.5% was calculated for 170,000 plants over an area of 3.5 hectares.

Studies have shown that it is feasible to count individual plants using ready-made drone-based products, and with machine vision/learning algorithms, image data can be converted into useful information for non-experts. Pros: The work is very good. I have done a lot of field experiments and there are actual data. Logical sound. Disadvantages: The method uses AlexNet and transfer learning, which is relatively general. Reference value: 3.5 points.

12. EfficientNet-B4-Ranger: A novel method for greenhouse cucumber disease recognition under natural complex environment

Abstract: In this study, our main task is to find an effective way to solve the problem of disease similarity caused by the influence of two diseases occurring in the same leaf and the influence of external light. First, we obtained a dataset of cucumber leaf diseases in a naturally complex greenhouse context, which included not only powdery mildew, downy mildew, healthy leaves, but also a combination of powdery mildew and downy mildew. Secondly, we used the current state-of-the-art method EfficientNet to build a classification model for the above four types with a model accuracy of 97%, and proved that EfficientNet-B4 is the most suitable method for this study. Finally, we constructed two classification models for cucumber-like diseases using the improved Effi cientNet-B4, a state-of-the-art optimization program, Ranger, with surprising accuracy (96%).

Methods: CNN. Based on the above related work, we chose the EfficientNet model as the method for disease classification research, and selected the current typical learning methods such as deep VGG, ResNet, AlexNet, Ierception v4, SqueezeNet, DensenNet comparison model for classification study of similar diseases cu-cumber leaves to participate in this study.

The purpose of this study is to:

1. To explore effective classification methods for greenhouse cucumber diseases (mainly including PM, DM, PD and healthy leaves) in natural and complex environments.

2. The current state-of-the-art EfficientNet model was improved by using the current state-of-the-art optimizer (Ranger), and applied it to the re-understanding of DM and PD cucumber diseases with high similarity.

3. Discuss the challenges and opportunities that may arise in the future for the identification and classification of plant diseases.

Dataset source: More than 5,000 samples of 4 types of pests and diseases were taken in one day. The total number of datasets was increased to 20,000 through reinforcement learning. The ratio of training set, validation set, and test set is 8:1:1.

Personal summary: This paper uses the EfficientNet model to classify mixed pests in cucumber leaves. The work can be done. Rationally structured. The method is so-so.

13.Automatic vegetable disease identification approach using individual lesion features

Summary: This study proposes a method to extract local features from a single greening and necrotic lesion, minimizing feature redundancy and vector size. The Color Consistency Vector (CCV), a feature that depicts different homogeneous patterns relative to disease progression, is extracted from the greened region. On the other hand, local binary patterns (LPBs) are extracted from necrotic areas. These individual lesion features are connected to form pathological feature vectors for disease identification, thereby minimizing feature size and avoiding the possibility of dealing with surface descriptors. To verify the effectiveness of the proposed method, we used different traditional classifiers (SVM, Naive Bayes, KNN) to test the quality and efficiency of these features in accurately classifying plant diseases. The results show that the method achieves high accuracy and recall rate in all cases, with a recall rate of more than 99%, which is improved compared with other methods reported in the literature.

The last three fully connected layers of AlexNet are pruned and replaced by new layers that will be divided into EB, LB, and HL levels. In this way, features from other layers are preserved, i.e., transferred layer weights. However, the weights and biases of the new layer have increased by a factor of 10, and the learning speed has become faster. ResNet is similar to the process applied in AlexNet transfer learning, where this layer and the output layer are replaced by a new layer, and the number of outputs is equal to the number of disease classes.

However, in this case, when the weights of the earlier layers are reinitialized, the weights of the first ten layers of the network are frozen by setting their learning rate to 0. This speeds up network training because the gradients in these layers are not updated. Conclusion: The use of a single greening and necrosis disease region pattern leads to the generation of effective pathological feature descriptors, minimizes the length of feature vectors, and improves identification results. In addition, with only 4 color features, the SVM classifier is able to achieve an overall accuracy of 99.33% with an AUC of 0.99. Similar results were obtained using only 7 subsets of the 15 feature descriptors out of a total of 15. This strongly emphasizes the importance of locating the diseased area as greenish and necrosis, as well as the importance of the color moment of the CCV in disease identification.

Personal summary: there is no innovation in the method, the structure is reasonable, and the logic is reasonable. reference

14. Comparison of convolution neural networks for smartphone image based real time classification of citrus leaf disease

Summary: Experts help farmers diagnose citrus diseases by using agricultural laboratories or observing visual symptoms. Due to the cost of specialists and the lack of laboratories, these methods may not be available to all farmers. In this study, two different convolutional neural network (CNN) structures were compared to classify citrus leaf diseases. In this paper, two CNN architectures, namely MobileNet and Self-Structured (SSCNN) classifiers, were used to detect and classify leaf diseases at the citrus plant stage.

A smartphone image based on a citrus disease dataset was prepared for the study. Both models were trained and tested on the same citrus dataset. The accuracy and loss of the training set and the validation set are used to evaluate the performance of the model. The optimal training accuracy of the MobileNet CNN is 98%, and the validation accuracy in epoch 10 is 92%. However, the best training accuracy of SSCNN in epoch 12 is 98%, and the verification accuracy is 99%. The results showed that the SSCNN algorithm based on smartphone images had high accuracy and practicability for citrus leaf disease classification. In addition, the SSCNN algorithm has a shorter computation time compared to MobileNet, which can be considered as a cost-effective method for citrus disease detection.

1.2 Semantic segmentation

1. Establishing a model to predict the single boll weight of cotton in northern Xinjiang by using high resolution UAV remote sensing data

Summary: Using the multi-temporal high-resolution visible light remote sensing data obtained by UAVs, a single bell weight prediction model is established. The remote sensing data of flowering and boll stages of 29 farmlands in Changji, Shihezi and Shawan areas of Xinjiang were studied. Five circular areas with a radius of 1 m were selected as the ground survey area in each field, and cotton boll samples were collected. As the dependent variable of the model, the full convolutional neural network was used to perform pixel-level semantic segmentation on the remote sensing image, extract the cotton segments in the image, and eliminate the influence of soil pixels on the accuracy of the model. The correlation analysis was performed by combining the VDVI (Visible Band Vegetation Index) at the flowering and boll setting stage, the VDVI cotton seed opening stage, the VDVI flocculation field (FCN extraction) and RGB values, and then using the least squares linear regression and BP neural network models to calculate the average single boll weight in the survey area using the above, middle, and low cotton layers.

Subsequently, K-fold cross-validation is performed to evaluate the results. The results show that the least-squares linear regression result (R2 = 0.8162) is almost equal to that of the BP neural network (R2 = 0.8170). The area of boll opening rate and VDVI at flowering and boll setting stage were highly correlated with the upper single boll weight. This study proposes a method to realize the large-scale prediction of single boll, which provides a new idea for cotton yield prediction and breeding screening.

Pros: There are field experiments and specialized equipment. Disadvantages: The amount of data is not large, and the workload is not large. It's the data that I got after running for a week with a drone. Then, the correlation coefficients between different parameters and the weight of a single bell were determined to determine the correlation between them, so as to achieve the purpose of predicting the output. Reference value: 3 points. , 1408 challenging leaf images were collected for this study. The results obtained in this study are novel for several reasons: (1) we have demonstrated that accurate automatic background removal of leaf images is possible using FCNNs field conditions, and (2) we have shown that semantic segmentation networks can be used to perform instance segmentation of an instance of an object to provide a target-dominant image even if the image contains other similar but less prominent objects.

2. Tomato leaf segmentation algorithms for mobile phone applications using deep learning

Abstract: In this paper, we propose a fully convolutional neural network to perform automatic background subtraction of leaf images captured in a mobile application. Used in a mobile app, the target leaf often dominates the image taken by the farmer. The leaves will also be surrounded by a variety of background features, including other leaves, stems, fruits, soil, and mulch.

The goal of segmenting the network is to remove these background features so that only the target leaves are retained. In order to train and test the proposed network, a dataset is prepared that represents this scenario. It consists of 1408 tomato leaf images. The proposed technique replaces competing background subtraction algorithms, but does not require user intervention and does not limit the orientation, shape, or illumination of the target blade. In addition, all CNN models are capable of segmenting a 256x256 pixel RGB image in 0.12 seconds when running on the GPU and 2.1 seconds when running on the CPU, which is much faster than any competing technology.

Conclusion: Our research has shown that semantic segmentation networks can be used to perform instance segmentation of an instance of an object to provide a target-dominant image even if the image contains other similar but less prominent objects.

In this paper, the performance of the two loss functions in the pixel classification layer is also evaluated. We have reported that GDL is superior to cross-entropy loss and performs cleaner mask boundaries. The proposed network implements a reserve test set of more than 0.96 mwIoU and 0.91 mBFScore. Specifically, we propose to design the KijaniNet test set with a score of 0.9766 mwIoU and 0.9439 mBFScore reserves, and we know that no other study in the existing literature yields similar results in leaf image segmentation tasks.

Summary: I personally didn't know how difficult this is, and it doesn't look like it's very difficult. But as far as the author describes, he seems to be the first person to do this. Reference value: 4 points.

3. Biophysical parameters of coffee crop estimated by UAV RGB images

Summary: This work aims to evaluate the accuracy of photogrammetry techniques using the SfM point cloud, which is used to estimate the height (h) and canopy diameter (d) of coffee trees from aerial images obtained by UAV with RGB (red, green, blue) cameras and compare the results with data from 144 trees measured in situ over 12 months. The rotary-wing drone is used in automatic flight mode and is connected to a conventional camera, flying at an altitude of 30m with an image overlap rate of 80% and a speed of 3m / s. Images were processed using PhotoScan software and analyzed in Qgis. The correlation between the h-value obtained in the field and the h-value obtained by the drone was 87%, and the correlation between the d-value obtained in the field and the value obtained by the drone was 95%. Using UAV–SfM images acquired with RGB digital cameras, a large number of estimates of attributes such as h and d of coffee trees can be obtained. Pros: There is a year of data. Disadvantages: The assessment and prediction of the height and diameter of his coffee tree cannot be summarized as such a wide range of physiological data. The direction of analysis used is correlation analysis. Reference value: 2 points.

4. Vineyard yield estimation by combining remote sensing, computer vision and artificial neural network techniques

Summary: The vitality of plants is measured by incorporating the vegetation index (VI) to perceive the health status of crops and predicting the final yield by obtaining vegetation fractional coverage (Fc) through computer vision. Multispectral images obtained from unmanned aerial vehicles (UAVs) can be used to acquire VI and Fc and be used with artificial neural networks (ANNs) to model the relationship between VI, Fc and yield. The proposed method was applied in a vineyard where different irrigation and fertilization dosages were used. The results show that the use of computer vision technology to distinguish between canopy and soil is necessary to obtain accurate results in precision viticulture. In addition, the combined use of VIs (reflectometry) and Fc (geometry) to predict vineyard yields resulted in higher accuracy (root mean square error (RMSE) = 0.9 kg vine-1 and relative error (RE) = 21.8%) compared to close to the use of VI (RMSE = 1.2 kg vine-1 and RE = 28.7%). The implementation of machine learning techniques yielded more accurate results than linear models (RMSE = 0.5 kg vine-1 and RE = 12.1%).

Advantages: The study had multiple processed data from a one-year physiological experiment, a large number of field experiments and controlled experiments to test the results of the images taken by the drone in predicting grape yield. The logic is clear and the structure is reasonable. Disadvantages: The method used is really not new in terms of computers. Reference value: 4 points.

5. Accuracy of carrot yield forecasting using proximal hyperspectral and satellite multispectral data

Summary: Proximal and distal sensors have demonstrated their effectiveness in estimating a number of biophysical and biochemical variables for many different crops, including yields. Evaluation of their accuracy in vegetable crops is limited.

In this study, we investigated the predictive accuracy of proximal hyperspectral and satellite multispectral sensors (Sentinel-2 and WorldView-3) for carrot root yield under different crop configurations, seasons, and soil conditions. Aboveground biomass (AGB), canopy reflectance and corresponding yield measurements were collected from 414 sites in 24 fields in Western Australia (WA), Queensland (Qld) and Tasmania (Tas). The best sensor (hyperspectral or multispectral) was determined with the highest overall determination coefficient between yield and different vegetation index (VIs) and linear and nonlinear models (linear and nonlinear models) as the best sensor ( VIs) and their impact on spatial resolution. The best regression fit for each region was used to extrapolate point source measurements to all pixels of each sampled crop, resulting in a predicted yield plot and estimating the average carrot root yield (t/ha) at the crop level.

The latter is compared with the yield (t/ha) of commercial carrot roots obtained from growers to determine the accuracy of the forecast. The measured yield varies between 17 ~ 113 t/ha, and the overall accuracy (% error) of the average yield prediction is 9.2% in WA, 10.2% in Qld and 12.7% in Tas. The yield correlation coefficient (R2 < 0.1) produced by VIs from hyperspectral sensors was lower than that of similar measurements from multispectral sensors (R2 < 0.57, p < 0.05). The spatial resolution was increased from 10 m to 1.2 m, and the regression performance was improved by 69%. It is not possible to make a non-destructive estimate of the spatial yield change of root vegetables such as carrots before harvest. The regression coefficient ranged from 0.27 to 0.77 and the error reached 1%. Advantages: There are abundant field experiments, and fields from different regions are used for comprehensive analysis to judge the accuracy of prediction. Cons: The actual effect is not good. Reference value: 3.5 points.

6. Evaluation of cotton emergence using UAV-based imagery and deep learning

Summary: The purpose of this study is to develop a novel UAV image processing method to achieve near-real-time processing of UAV images. In this study, an unmanned aerial vehicle (UAV) imaging system was used to collect RGB image frames of cotton seedlings to evaluate the number of stands and canopy size. The image is pre-processed to correct for distortion, and the ground sample distance and the number of georeferenced cotton rows in the image are calculated.

A pre-trained deep learning model, resnet 18, was used to estimate the number of stands and canopy size of cotton seedlings in each image frame. The results show that the method can accurately estimate the forest fraction on the experimental dataset, R2 = 0.95. Similar results were obtained in the test dataset with an estimated accuracy of R2 = 0.93 for canopy size. The processing time of a 2000 M pixel image frame with a georeference for each crop row is 2.22 s (including 1.80 s for preprocessing), which is more efficient than traditional mosaic-based image processing methods. An open-source automated image processing framework was developed for cotton emergence-assessment and could provide effective data processing and analysis to the community.

In this study, we developed an effective image data processing and analysis framework for timely assessment of cotton emergence using UAV-based RGB images and deep learning technology. This study differs from previous published studies in that the approach is to directly process individual image frames, rather than developing orthomorphic images to reduce processing time. Specific goals include: (1) developing a pre-processing pipeline to segment and georeference crop rows in each individual image frame; (2) implement a deep learning model to estimate cotton quantity and canopy size; (3) Establish a framework for automatic generation of cotton emergency maps based on geographic information. Pros: Lots of field experiments and drone monitoring pictures. Cons: Don't know. Reference value: 2 points. Because we don't have the equipment and we don't have the experimental grounds.

7. Deep learning techniques for automatic butterfly segmentation in ecological images

Abstract:Deep learning-based methods are more promising than traditional methods in butterfly ecological image segmentation, because they have strong feature learning and representation capabilities. However, butterfly segmentation is still challenging when complex background interference occurs in the image.

To solve this problem, we propose an extended encoder network to capture more advanced functions and obtain a high-resolution output that is both lightweight and accurate for automatic butterfly ecological image segmentation. In addition, we use the dice coefficient loss function to better balance the butterfly and non-butterfly regions.

Conclusion: In this paper, an expansive autoencoder network is proposed. First, the original butterfly ecological image is input into the dilated feature encoder module (dilated convolution). The extracted features are then fed into the feature decoder module and a segmentation map is generated. In addition, we use the dice coefficient loss function to better balance the butterfly and non-butterfly regions. Experimental results on the Leeds butterfly dataset show that our method is superior to the image segmentation method based on the latest deep learning. Experimental results show that the proposed method basically overcomes the problem of automatic segmentation of butterflies in ecological images.

The main contributions of this work are summarized as follows: (1) we propose an HCDC module to capture more advanced functions and obtain high-resolution outputs; (2) we integrate the proposed HCDC block with the encoder-decoder structure to construct a novel network structure called "Extended Encoder Network" (DE-Net); (3) To the best of our knowledge, this is the first attempt to segment complex butterfly ecological images using deep learning. In our experiments, five state-of-the-art deep learning methods designed for natural image segmentation were applied to the Leeds butterfly dataset to compare with the proposed methods. Experimental results show that the accuracy of the proposed method is better than that of the latest method. (4) The proposed DE-Net is 1.2 to 6 times smaller than the butterfly ecological image segmentation problem, which can achieve better learning ability than the large model, thereby saving storage space and improving the portability of the model.

Personal Summary: This paper uses a network dataset. Methods of semantic segmentation. I don't know if the method is innovative. The structure is reasonable, the logic is reasonable. Reference value: 4 points.

8. Improving segmentation accuracy for ears of winter wheat at flowering stage by semantic segmentation Semantic segmentation was used to improve the segmentation accuracy of winter wheat ears at flowering stage

Abstract: In this study, a semantic segmentation-based method, EarSegNet, is proposed to perform pixel-level classification to segment wheat ears from canopy images captured under field conditions. EarSegNet integrates an encoder-decoder structure and extended convolution to further improve the segmentation accuracy and efficiency of winter wheat ears. The results showed that the proposed EarSegNet was able to achieve accurate segmentation of wheat ears from canopy images captured at flowering (segmentation quality = 0.7743, F1 score = 87.25%, structural similarity = 0.8773).

To validate the proposed method, the performance of the proposed EarSegNet was compared with the widely used segmentation methods, i.e., SegNet, two-stage method, and Panicle-SEG. The results show that the proposed EarSegNet is superior to the compared methods, making it a powerful and effective tool for segmenting the ears of winter wheat based on the canopy images captured at the sowing stage. Generalization tests show that the performance of the proposed EarSegNet is better than that of the comparison method, indicating that the EarSegNet has great potential for field applications. The results obtained show that the depth of the encoder (i.e., VGG16) has no significant effect on the performance of EarSegNet, however, deepening VGG16 can improve the performance of the recall evaluation metric of EarSegNet. Methods: 36 original images were obtained for each growing season.

The images had a pixel resolution of 5184 × 3456, then were manually cropped and reshaped to 2500 × 2500 pixel resolution. Since each cell has 3 ROI images, 2 ROI images are randomly selected from each cell to build a semantic segmentation network, and the left one is used for performance evaluation, that is, 48 images are used to build the model and 24 images are used for performance evaluation. Eventually, the number of images was expanded to 3,000. Personal summary: The structure is clear and logical. It seems to be a little innovative. Reference value: 4 points.

2.3 Object Detection

1. Brown rice planthopper (Nilaparvata lugens Stal) detection based on deep learning

Summary: Brown planthoppers are one of the main pests of rice. Rapid and accurate detection of rice planthoppers helps to deal with rice in a timely manner. Due to the small size, large number and complex background of BRPHs, it is challenging to detect BRPHs. This paper proposes a two-layer detection algorithm based on deep learning technology to detect them. The algorithm for these two layers is a faster RCNN (region with CNN features). In order to make efficient use of computing resources, different feature extraction networks were selected for each layer. In addition, the Layer 2 detection network was optimized to improve the final detection performance.

The detection results of the two-layer detection algorithm were compared with the detection results of the single-layer detection algorithm. The detection results of the double-layer detection algorithm for detecting different populations and numbers of BRPHs were tested and compared with the deep learning object detection network YOLO v3. The test results show that the detection results of the two-layer detection algorithm are significantly better than those of the single-layer detection algorithm. In different numbers of BRPHs tests, the average recall rate of the algorithm is 81.92%, and the average accuracy rate is 94.64%. The average recall rate of YOLO v3 was 57.12% and the average correct rate was 97.36%. In the BRPHs experiments of different ages, the average recall rate of the algorithm was 87.67%, and the average accuracy rate was 92.92%. The average recall rate of YOLO v3 was 49.60% and the average accuracy rate was 96.48%. Advantages: reasonable structure. Disadvantages: The ratio of the training set, the test set, and the validation set is 8:1:1. The dataset is smaller. Reference value: 1 point.

2. Identification of olive fruit, in intensive olive orchards, by means of its morphological structure using convolutional neural networks

Summary: Due to the high economic value of olive cultivation, accurate yield estimates are an important goal of olive cultivation. This paper proposes a methodology that aims to achieve this purpose. It includes an artificial vision algorithm capable of detecting a digital image of the visible olive tree, captured at night, directly in the field under artificial lighting. The photos were taken in September 2018 (two months before the harvest) at Picual Olea europaea L. a dense olive orchard. In this method, the images are first preprocessed to generate a set of high-probability sub-images containing olives, thereby narrowing the search space to the order of 103. Next, these sub-images are classified by Convolutional Neural Networks (CNNs) as olives, if they are centered around olive fruits, or in any other case (even if they contain peripheral fruits).

To train and validate the CNN, a special database called OLIVEnet is compiled into two disjoint sets that integrate these subimages. The training set and validation set were constructed with 234, 168 and 299,946 olive sub-images and other sub-images, respectively. We tested five different CNN topologies and correctly classified the best performing one in 83.13% of olive instances with 84.80% accuracy and 99.12% accuracy in other instances; The measurement accuracy is 0.9822 and the F1 score is 0.8396. To the extent of the author's knowledge, this paper proposes the first image analysis method to automatically identify olive fruit in images of whole trees taken directly in the field. Personal summary: This paper is not bad.

2. Machine learning

1. A random forest ranking approach to predict yield in maize with uav-based vegetation spectral indices

Summary: A sorting-based method is proposed to enhance the potential of RF method for maize yield prediction. The method is based on the relevant parameters of individual vegetation indices (VIs). VIs are ranked separately based on a value metric that measures improvement in Pearson's correlation coefficient by using the RF vs. baseline method. Therefore, only the most relevant VIs are considered to be the input features of the RF model. We used 33 VIs extracted from multispectral unmanned aerial vehicle (UAV) images.

Two different sensors, Sequoia and MicaSense, were used to generate multispectral data; The 2017/2018 and 2018/2019 crop seasons are respectively. NDVI, NDRE, and GNDVI ranked the top three among all evaluation indicators, and their combination with RF improved maize yield forecasts. The conclusion is that the sorting-based vegetation restoration strategy indices (VIs) can potentially realize the random forest (RF) algorithm to predict maize yield using only multispectral drone images. Pros: 2 years of data, images taken with drones. There are 11 varieties of maize and 33 related indicators. Large amount of data. Disadvantages: The results of the Daejeon experiment data do not appear in the text. And the operation method is quite simple.

3. Miscellaneous

1. DropLeaf: A precision farming smartphone tool for real-time quantification of pesticide application coverage

Summary: This work introduces and experimentally evaluates a novel tool that can be used as a smartphone-based mobile application called the DropLeaf-Spraying Meter. Tests conducted with DropLeaf have shown that pesticide coverage can be estimated with high accuracy, despite the ease of operation. Our method is based on the development of custom image analysis software for real-time evaluation of spray deposition of water-sensitive paper. The proposed tool can be widely used by farmers and agronomists who carry conventional smartphones, and DropLeaf can be easily used for spray drift assessment of a variety of methods, including emerging drones and smart sprayers. This paper talks about a developed app called DropLeaf. It can be used to assess pesticide spray coverage. Pros: Developed a usable APP. Cons: There are similar apps on the market. Reference value: 2 points.

2. Underutilised crops database for supporting agricultural diversification

Summary: This article builds a usable global access database in an attempt to store information for underutilized crops. The focus on designing relevant agricultural databases, data standards and crop diversification was examined, and a data model was developed that included the main elements of the crop value chain in the food system. To guarantee the accuracy of the data, we added a metadata table that stores information about all the data sources recorded in the database. and built a web-based user interface for data management and access. The open-access user interface allows for simple data sorting and filtering operations based on user needs. Pros: A database was set up, a website was developed. Cons: Don't know. Reference index: 2 points.

3. Replacing traditional light measurement with LiDAR based methods in orchards

Summary: In this paper, the virtual model of trees can be used to analyze the lighting environment and geometric measurements such as canopy volume. However, these simulation models allow for the analysis of variables that cannot be directly measured. The study presented a study of a lidar-based approach to replace light interception using a light vehicle, showing good results with a wide range of applicability. (This paper is to investigate the effect of LiDAR lidar in capturing optical data and compare it with the results of traditional methods.) Advantages: A new method to measure the light interception efficiency of trees was explored and its possibilities were verified. Cons: Don't know. Reference index: 3 points.

4. A cyber-physical intelligent agent for irrigation scheduling in horticultural crops

Abstract:This paper introduces the design and implementation of an electrophysical crop irrigation system based on the concept of intelligent agent. The system allows the acquisition of field information via sensors, the addition of water and the activation of solenoid valves according to the decision-making system. Communication Capability and Performance Measurements, Environmental, Actuation, and Sensing Systems (PEAS) are explained. With a central station and its internet connection, monitoring can be carried out remotely in the field or from other locations. The performance evaluation of the system was carried out through the use of crop modeling software and a pilot crop in the large-scale irrigation and drainage areas of Boiacá and Ferravitoboba, Colombia.

The developed system can keep soil moisture close to the water holding capacity of the field through several irrigation strategies and avoid water wastage and overuse. The innovation of the paper lies in the integration of the agent's reasoning and active abilities into the embedded board. The developed system allows the use of an embedded system as a central station for irrigation scheduling and water management, with a centralized server for information storage and monitoring, using equipment with internet access.

Methodology: The cyber-physical design principle follows the concept of an intelligent agent based on the integration of multiple subsystems, as shown in Figure 1. The agent consists of an intelligent multi-sensor array (SMSA) located in the field and a weather station subsystem (WSS) that completes the functions of the perception system. In addition, agents use the Irrigation Activation Subsystem (IAS) to perform operations in the field.

The radio transceiver module is used to generate a mesh network between the SMSA, the IAS, and the Agent Central Station (ACS) located in the farm. A web server was implemented to fetch data from de ACS and WSS. Other systems can refer to the web server database to monitor task support activities in decision-making. The SMSA and IAS subsystems use ATmega328p microcontrollers. The WS and ACS were developed using a Raspberry-Pi-3b® board.

Fig. 1. Cyber-physical intelligent agent architecture for irrigation scheduling

Fig. 2. Smart Multi-Sensor Array (SMSA). Source: Authors.

Fig. 3. Irrigation Activation Subsystem (IAS). Source: Authors.

In a soil sensor formulation, depletion and watering time are determined based on measurements from the soil moisture sensor. These results confirm that the intelligent agent irrigation scheduling system based on yield and water use efficiency prediction is suitable for irrigation scheduling and can improve the water-saving effect.

Summary: This paper mainly talks about how to use sensors and some hardware configurations to determine the water demand for crop irrigation to achieve the purpose of water conservation. Pros: There is hardware and experiments. Cons: Don't know. Reference value: 4 points.

5. Biophysical parameters of coffee crop estimated by UAV RGB images

The development of digital agriculture to estimate the biophysical parameters of coffee crops using UAV RGB images combined with computational tools and unmanned aerial vehicles (UAVs) has enabled the collection of data to reliably extract vegetation indices and biophysical parameters derived from kinematic structure (SfM) algorithms. The aim of this work was to evaluate the accuracy of photogrammetry using SfM point clouds in estimating coffee tree height (h) and crown diameter (d) in aerial images acquired by UAVs using RGB (red, green, blue) cameras, and to compare the results with 12 months of field measurement data. The experiment was conducted on a coffee plantation in Minas Gerais, Brazil. The rotary-wing UAV adopts the autonomous flight mode, coupled with the conventional camera, flies at an altitude of 30m, the image overlap rate is 80%, the speed is 3m/s, the image processing adopts PhotoScan software, and the analysis is carried out in Qgis.

The correlation between the h-value measured in the field and the h-value measured by the drone was 87%, and the correlation between the d-value measured in the field and the d-value measured by the drone was 95%. UAV–SfM images acquired with RGB digital cameras allow important estimates of attributes, such as h and d of coffee trees.

6. A satellite-based ex post analysis of water management in a blueberry orchard

Abstract: In the current context of water scarcity caused by climate change and increased demand for water for food production, farmers must adapt to water management and shift from supply-driven to demand-driven water management, taking into account trade-offs among stakeholders. Quality, quantity and cost. Therefore, agricultural practices must take full advantage of technology, research and development, and adapt to local requirements. Today, remote sensing is a useful tool for estimating crop water requirements (evapotranspiration) and for mapping their spatial and temporal variations. In this work, we propose a new method that allows users to conduct an audit (ex post facto) of irrigation strategies in blueberry fields in central Chile using an irrigation decision support system called AquaSat® as the primary tool.

The tool combines satellite information with field data and provides information on the spatial distribution of water used by crops to manage farm-scale irrigation. The main contribution of this work is to detail a new approach to irrigation management by comparing the amount of water applied, evapotranspiration and potential demand. The program allows users to audit current irrigation management and determine the impact on productivity. Based on our results, we can conclude that the entire irrigation sector did not use enough water to apply to the farm in two seasons to reach the potential blueberry yield.

In this work, we propose a new approach that allows the use of an irrigation decision support system called AquaSat® to review (post-mortem) irrigation strategies. The system combines satellite information with field data and provides a spatial and temporal distribution of crop evapotranspiration (ETc) and compares it to water application and potential demand (water to reach maximum yield). This paper uses the example of blueberry irrigation in central Chile to illustrate the potential of this novel approach. In addition, AquaSat® is a decision support system developed by the authors' team which allows to estimate crop evapotranspiration (etc.) and their water requirements, according to irrigation projects. It is also a useful tool for reviewing farmers' irrigation strategies after the fact (Lillo-Saavedra, 2019).

Personal summary: The logic of the method is clear. This paper feels very well done, experiments and model design. But I still don't know how to measure the evaporation of leaves, through satellite information.

7.A 3D white referencing method for soybean leaves based on fusion of hyperspectral images and 3D point clouds

Abstract: Due to the non-uniform light intensity of the light source at different wavelengths, the original image needs to be calibrated using a white reference. The flat white panel is usually scanned as a white reference. However, geometric factors such as the tilt angle of the blade cannot be calibrated with a flat white datum. In this publication, the effectiveness of calibrating the corresponding raw image with an angled white reference was demonstrated for the first time. In addition, a 3D white reference library was created that integrates the hyperspectral camera and the Kinect V2 depth sensor system with different angular and spatial positions.

In this way, the pixels on the blade surface can be calibrated by the closest point to the tilt angle and spatial position in the 3D reference library. The validation sample for this reference library is greenhouse soybean leaves. The results show that compared with the traditional plain white reference calibration, the reflectance spectrum after the three-dimensional calibration is closer to the standard calibration (flat white reference calibration). In addition, the distribution of pixel-level normalized difference vegetation index (NDVI) on the surface of soybean leaves after three-dimensional calibration was closer to the standard calibration. This method has the potential to improve the quality of plant image indexing. Combined with lidar sensors, this new approach has the potential to be applied in the field. Keywords: plant phenotype, inhomogeneous illumination, hyperspectral image calibration, 3D white reference library, 3D point cloud. I don't understand this paper very well.

8. Assessing winter wheat foliage disease severity using aerial imageryacquired from small Unmanned Aerial Vehicle (UAV)

Abstract: Red, green, and blue band (RGB) images were acquired by flying a rotorcraft. The images were then processed to develop a positive mosaic, and three vegetation indices were calculated. The resulting image dataset is further processed to generate plot-level data. Visual records of field reactions and leaf rust severity were recorded to calculate the infection coefficient (CI). During these two years, significant differences in vegetation indices were found between wheat genotypes. The Standardized Difference Index (NDI), Green Index (GI) and Green Leaf Index (GLI) were linearly correlated with CI, with R2 values ranging from 0.72 to 0.79 (p < 0.05) in 2017 compared to R2 values ranging from 0.63 to 0.68 in 2017 (p < 0.05) in 2018. The Ground Normalized Difference Vegetation Index (NDVI) also showed a significant correlation with CI in both years (R2 = 0.86, p < 0.05, 2017; R2 = 0.83, p <0.05, 2018). The results suggest that drone imaging and automated data extraction can facilitate the acquisition of high-throughput phenotypic data for disease severity grades.