Detecting Overgrown Plant Species Occluding Other Species in

Complex Vegetation in Agricultural Fields Based on Temporal

Changes in RGB Images and Deep Learning

Haruka Ide

, Hiroyuki Ogata

, Takuya Otani

, Atsuo Takanishi

and Jun Ohya

Department of Modern Mechanical Engineering, Waseda University, 3-4-1, Ookubo, Shinjuku, Tokyo, Japan

Faculty of Science and Technology, Seikei University, 3-3-1, Kichijoji-kitamachi, Musashino-shi, Tokyo, Japan

Waseda Research Institute for Science and Engineering, Waseda University, 3-4-1, Ookubo, Shinjuku, Tokyo, Japan

Keywords: Synecoculture, Agriculture, Deep Learning, Robot Vision.

Abstract: Synecoculture cultivates useful plants while expanding biodiversity in farmland, but the complexity of its

management requires the establishment of new automated systems for management. In particular, pruning

overgrown dominant species that lead to reduced diversity is an important task. This paper proposes a method

for detecting overgrown plant species occluding other species from the camera fixed in a Synecoculture farm.

The camera acquires time series images once a week soon after seeding. Then, a deep learning based semantic

segmentation is applied to each of the weekly images. The plant species map, which consist of multiple layers

corresponding to the segmented species, is created by storing the number of the existence of that plant species

over weeks at each pixel in that layer. Finally, we combine the semantic segmentation results with the earlier

plant species map so that occluding overgrown species and occluded species are detected. As a result of

conducting experiments using six sets of time series images acquired over six weeks, (1) UNet-Resnet101 is

most accurate for semantic segmentation, (2) Using both segmentation and plant species map achieves

significantly higher segmentation accuracies than without plant species map, (3) Overgrown, occluding

species and occluded species are successfully detected.

1 INTRODUCTION

In the field of agriculture, with modernization, a

cropping method characterized by the cultivation of

specific plants as monocultures and the use of

chemical fertilizers and pesticides has been adopted

to enhance food productivity (Tudi et al., 2021).

However, such conventional farming practices render

plants vulnerable to pests, diseases, and weeds.

Furthermore, the continuous and increasing use of

chemicals not only disrupts the soil ecosystem but

also reduces the biodiversity of agricultural land and

its surrounding environment (Conway and Barbie,

1988; Geiger et al., 2010; Savci, 2012). Particularly,

reducing biodiversity is severe in conventional

farming, prompting the need for achieving more

sustainable farming methods (Norris, 2008).

In response to these issues, a new method called

“Synecoculture” has been proposed. Synecoculture is

an agricultural method in which various plant species

are grown in mixed, densely planted environments in

a single farm to promote self-organization of the

ecosystem, thereby increasing the biodiversity of the

farm and producing useful crops by enhancing

ecosystem functions. On farms where Synecoculture

is practiced, the rich biodiversity results in intensive

competition for survival among plant species. It is

believed that plants' inherent self-organization ability

allows the plants to grow without using pesticides or

chemical fertilizers. Therefore, Synecoculture is

considered to be a more sustainable agricultural

practice than conventional agriculture because

Synecoculture can increase the productive capacity of

multiple crops as the diversity of plant populations

expands.

Synecoculture is also expected to convert deserts

and wild areas into green spaces in the future. In such

cases, Synecoculture plantations could have very

large areas, and automation of their management is

essential. However, the vegetation in Synecoculture

plantations is so complex as mentioned earlier that it

is difficult to automate the management using

conventional agricultural machineries.

266

Ide, H., Ogata, H., Otani, T., Takanishi, A. and Ohya, J.

Detecting Overgrown Plant Species Occluding Other Species in Complex Vegetation in Agricultural Fields Based on Temporal Changes in RGB Images and Deep Learning.

DOI: 10.5220/0012352500003654

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 266-273

ISBN: 978-989-758-684-2; ISSN: 2184-4313

The third and fourth coauthors of this paper focus

on the development of an automation robot designed

for the management of Synecoculture (Tanaka et al.,

2022). The 1st, 2nd and 5th coauthors are specifically

investigating the visual capabilities of this robot. Our

objective is to automate crucial tasks such as pruning

overgrown dominant plants, seeding, and harvesting

crops. Especially in Synecoculture farms, it is

important to increase plant diversity and the soil

should not be exposed. Therefore, it is not always

right to prune all the dominant species that are

commonly referred to as weeds. On the other hand, if

these dominant species are not pruned at all, only

these species will thrive, and they will overgrow and

inhibit the growth of other plants.

In order to prune dominant species that are

overgrown, occluding other species, and reduce

diversity, we have been working on a method using

image processing technologies that segment

dominant plants in the image acquired by the camera

observing Synecoculture farms and estimate pruning

points in a densely overgrown plantation. However,

due to its occlusive environments, it is difficult to

accurately segment dominant species and estimate

pruning points in agricultural areas with densely

mixed vegetation (Ide et al., 2022). It tunred out to us

that it is difficult to identify vegetation from an image

after plants have thrived.

The method proposed by this paper utlizes the fact

that there is little occlusion between plants during

short duration after seeding and the segmentation

accuracy for each plant is high. Our proposed method

acquires time series images using a fixed camera

according to some constant interval starting from a

time instance soon after the seeding. Each image of

the time series images are segmented, and plant

species map count the number of appearance of each

plant species at each pixel in the segmented images

over the time. By using the plant species map and

segmented time series images, even if useful species

are covered with other overgrown species,

information such as occluding areas, which should be

pruned, can be obtained.

This paper demonstrates that the utilization of

time-series data enables a more accurate

segmentation of the vegetation in Synecoculture

environments. Consequently, we have developed the

capability to identify plants that cause occlusion,

which is helpful to maintain the Synecoculture farm

automatically.

2 RELATED WORKS

Among agricultural robots already in practical use,

Raja et al. detected and pruned weeds that inhibited

the growth of useful species without pruning them

(Raja et al., 2020). They marked tomato and lettuce

stems, removing 83% of weeds without pruning.

However, due to the intricate and densely mixed

vegetation in this study, controlling specific plant

species through marking is deemed impossible.

In the field of machine learning, many studies

have been conducted to estimate plant species. In

recent years, deep learning has made it possible to

recognize a wide variety of plant species. Mortensen

et al. reported that the semantic segmentation model

achieved 79% accuracy, showing the effectiveness of

data augmentation and fine-tuning in plant species

estimation (Mortensen et al., 2016). Picon et al. used

Dual-PSPNet to classify diverse plant species,

achieving higher accuracy than the original PSPNet

model (Picon et al., 2022).

However, in all studies, recognition accuracy was

lowered for plants under the presence of occlusion,

which indicates that the presence of occlusion is a

major issue in classifying plant species using images.

To tackle the occlusion problem, Yu et al. improved

recognition accuracy over existing learning models

by developing original learning models suitable for

the presence of occlusion (Yu et al., 2022). However,

in Synecoculture farms, the sizes of plants vary due

to individual plant growth, which results in more

complex occlusions.

Therefore, it is necessary to enable highly

accurate plant species recognition in Synecoculture

environments, which could yield severe occlusions.

More specifically, when pruning, it is important to be

able to identify not only the plant species that are

visible from the camera, but also the plants that are

covered underneath.

3 PROPOSED METHOD

In Synecoculture farms, despite occlusions, it is

crucial to grasp the vegetation and identify occluded

useful species and occluding dominant species.

Our method leverages the initial stages of

Synecoculture farms, where occlusion is infrequent

right after seeding and germination. In these early

phases, small and isolated individual plants allow

accurate segmentation with existing models. As

shown in Figure 1, Our proposed method consists of

acquiring time series images, semantic segmentation,

Detecting Overgrown Plant Species Occluding Other Species in Complex Vegetation in Agricultural Fields Based on Temporal Changes in

RGB Images and Deep Learning

267

generating the plant species map and detecting

overgrown plant species occluding other species.

Figure 1: Over-view of our proposed method.

3.1 Semantic Segmentation

3.1.1 Model Architecture

UNet and Deeplab V3+ have been used as the main

learning models in many studies using semantic

sesgmentation with deep learning on plants (Li et al.,

2022; Wang et al., 2021; Zou et al., 2021; Kolhar and

Jagtap, 2023). In these studies, UNet and Deeplab

V3+ tend to perform better than other basic learning

models such as PSPNet and SegNet. Therefore, as a

result of comparing UNet and Deeplab V3+ in

Section 4.1.2, this paper uses Resnet101-UNet for the

subsequent processes such as for creating plant

species map.

Resnet50 and Resnet101 are used as the backbone

of the encoder part of the system. Recently, other

studies have shown that they are effective as the

backbone of semantic segmentation models

(Bendiabdallah et al., 2021; Nasiri et al., 2022;

Sharifzadeh et al., 2020).

Many studies have shown that fine-tuning of a

pre-trained model with a large dataset using a small

number of their own images is effective; therefore, we

use a backbone that was previously trained by

ImageNet and trained it in the form of fine-tuning.

In this paper, Section 4.1 investigate learning

models suitable for plant species detection in a

Synecoculture environment by changing the

combination of these encoders and decoders.

3.1.2 Implementation of Semantic

Segmentation

In this process, we obtain the semantic segmentation

result from each time series image. Before predicting

the species at each pixel, the original 1280 x 720

pixels image was cropped into 475 square pixels

before training models and prediction to reduce the

burden of computers. To prevent a decrease in

accuracy at the edges of cropped images, we used the

following method for image cropping. First, we

applied 400 pixels of zero padding to the outside of a

1280 x 720 pixels image and cropped the image to

475 x 475 pixels from the edge. When cropping the

image to 475 square pixels from the edge, we set the

strides to 100 pixels to generate overlaps between the

crop images. Each crop image outputs a semantic

segmentation result with a pre-trained model. Since

the overlapping regions of multi crop images produce

multiple predictions for a single pixel, the most

frequently predicted class was used as the prediction

class for that pixel. By outputting semantic

segmentation results as described above, a more

accurate 1280 x 720 pixels prediction result can be

obtained.

3.2 Creating Plant Species Map

To improve the accuracy of plant species estimation

under severe occlusions, the plant species map for

each species is generated using the plant species

estimation results from time series images. Figure 2

overviews how to create the plant species map.

First, the field is divided into several (six in this

paper) areas. A fixed bird’s eye view camera observes

each area as explained in Section 4.2 and acquires

RGB images at some predefined constant interval.

Next, for each of the acquired time series images,

we estimate the plant species at each pixel using a

semantic segmentation model that has already been

trained using training images as explained in Section

3.1.2. The output pixels by pixel-wise plant species

semantic segmentation results are then layered for

each plant species, as shown in Figure 2, where each

pixel of each plant species map stores a value

indicating whether or not that plant species is present:

specifically, 1 if the corresponding plant species is

present, and 0 otherwise.

Finally, the value of each corresponding pixel in

all the plant species layers are added over all of the

time series images so that the final plant species map

is obtained. This allows the presence of a plant to be

inferred from the plant map even under difficult-to-

recognize situations such as severe occlusions. That

is, even if a particular plant species is covered by

other plant species from some time instance, the

occluded plant species map is recorded as non-zero

pixels values, which means that the plant species map

is robust against occlusions.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

268

Figure 2: The method of creating plant species map.

3.3 Occlusion Area Detection

Even if a plant species is not detected in the

segmentation result of the image acquired at some

time instance, if that species is detected in the

segmentation result of an earlier image, we can know

the existence of that species from the plant species

map. Therefore, if a plant is not detected in the

segmentation result of the image at some time

instance and is confirmed that plant’s existence in the

plant species map, then that species can be judged to

be occluded by other species from an earlier time

instance.

The specific procedure for judging whether some

plant species overgrows from a time instance,

occluding other species, is as follows.

As a general case, we explain the case in which

the n-th week’s overgrowth is to be detected, where 2

≤ n ≤ 6. First, in the (n - 1)-th week’s plant species

map, whether the useful species (in this paper, green

pepper) exists at each pixel is checked. Next, in the

n-th week’s segmentation result, whether the useful

species exists is checked at the pixels whose image

coordinates are same as those of the (n - 1)-th week’s

plant species map’s pixels at which the useful species

exists. Finally, if all of the following three conditions

(1) to (3) are satisfied at a pixel, our proposed method

judges that a dominant species (apple mint or green

foxtail) overgrows at the pixel at the n-th week,

occluding the useful species.

(1) At a pixel in the (n - 1)-th week’s plant species

map, the useful species exists.

(2) At the n-th week’s segmentation result’s pixel

that corresponds to the pixel described in (1), the

useful species does not exit.

(3) At the n-th week’s segmentation result’s pixel

described in (2), a dominant species exists.

The above-mentioned processes are repeated for 2≤ n

≤ 6.

As our future work, the obtained positional

information on the overgrown, occluding non-useful

species could be utilized for determining the positions

to be pruned by our robotic system.

4 EXPERIMENTS AND RESULTS

4.1 Semantic Segmentation Model

As explained in Section 3.1 we compared the

common semantic segmentation models to chose

appropriate deep learning model for our data.

4.1.1 Training of Models

The image data used for training the models were

acquired once a week between May 2022 and June

2023 in farms in Akiruno, Tokyo, and Oiso,

Kanagawa Prefecture, where Synecoculture is

actually implemented. The data were acquired using

an Intel RealSense D435 RGBD camera, which

captured RGB images from a height of approximately

1.5 meters above the ground, looking vertically

downward at the farms.

A total of 137 1280 x 720 pixels images were

acquired, 127 were used to train the models and the

remaining 10 were used to test the models. The

acquired images were annotated so that "green

foxtail," "apple mint," "green pepper," "other plants,"

and "unvegetated area" were correctly labelled as

pixel wise segmentation. The train images were

cropped to 475 x 475 pixels before training.

Train images were further divided into train and

validation at a ratio of 7:3 for the training. In addition,

to compensate for the lack of datasets, data

Detecting Overgrown Plant Species Occluding Other Species in Complex Vegetation in Agricultural Fields Based on Temporal Changes in

RGB Images and Deep Learning

269

augmentation was applied to train images only. First,

we applied the gamma correction, which is

considered effective for plant detection (Saikawa et

al., 2019). In addition, we applied the Random

Shadow, Random Sun Flare, and Random Contrast

functions. Random Scale, Random Rotation, Color

Jitter, and Random Horizontal Flip, which are

commonly used, were also applied.

The GPU used for training was a 24GB NVIDIA

GeForce RTX 3090 with a learning rate of 0.001,

Adam as the optimization function, and a batch size

of 16. During the training process, mean IoU which is

explained in Section 4.1.2 is calculated, and the one

with the highest mean IoU value of validation during

each 200 epochs was used.

4.1.2 Model Performance Assessment

10 images of Synecoculture farms were used to test

the models, as described in section 4.1.1. To obtain

semantic segmentation results for each model, the

process described in section 3.1.2 was performed.

After the segmentation results for Deeplab V3+ or

UNet combined with Resnet50 or Resnet101 as

backbone were obtained, their accuracy was

calculated. Table 1 compares the accuracy of the

segmentation models on test images.

According to Table 1, overall UNet performs

better than DeeplabV3+. In particular, all the indices

of UNet with Resnet101 as the backbone are the

highest, while UNet with Renet50 almost the second

highest.

Table 1: Comparison of segmentation models.

Model

Recall

Precision

mIoU

Resnet50-

UNet

72.55 78.32 74.82 61.01

Resnet101-

UNet

77.63 79.19 78.02 64.58

Resnet50-

DeeplabV3+

73.38 76.24 74.48 60.40

Resnet50-

DeeplabV3+

73.49 76.09 74.18 59.93

4.2 Acquisition of Time Series Data

As explained in Section 3.2, time series images for

the plant species map were acquired.

For acquiring time series images, 1280 x 720

pixels RGB images were acquired once a week over

the six weeks from May 2023 using an Intel

RealSense D435 RGBD camera pointing vertically

downward from a bird's eye view fixture installed at

the top of the 110mm-high frame. The camera was

installed in the same position as that of the

agricultural robot under development. The farm is

divided into six areas, at each of which the camera

was placed, so that (= 6 times 6) time series images

of the Synecoculture environment were acquired for

the six areas over a total of six weeks. In addition, we

pruned the overgrown apple mint and green foxtail in

7-th week and acquired post-pruning images in the

same way as the method explained earlier. We also

applied black masks to equipment in the images that

had no relevance to the plants.

4.3 Segmentation Results in Time

Series Images

We then segmented the six images for each week

using the Resnet101-UNet trained model, applying

the method described in Section 3.1.2. Figure 3 shows

examples of the segmentation results obtained by the

Resnet101-UNet trained model in the four weeks out

of the six weeks. As shown in Figure 3, the green

pepper, which was completely visible in early weeks,

is covered by apple mint and green foxtail, which tend

to dominate after week 5, and only a part of the leaves

of the green pepper is visible due to the occlusion

caused by the overgrown apple mint and green

foxtail. To solve this problem, as described in Section

3.2, the plant species map is proposed by this paper.

Figure 3: Segmentation results for each week.

4.4 Creating Plant Species Map

Figure 4 compares segmentation results for each

single image with the plant species map for all over

the six weeks, where the color in the right three

columns indicates as follows: dark blue and bright

orange if the count at a pixel is small and large,

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

270

Figure 4: Visualization of plant species map.

Table 2: Comparison of segmentation accuracies with/without plant species map.

“with” / “without” Class (plant species) Precision % Recall % F1 % mean F1 %

“without” (Only using

week 6’s segmentation

result, without using plant

species map)

Green foxtail 83.49 46.18 59.47

46.68

Apple mint 97.51 56.17 71.28

Green pepper 88.43 4.90 9.28

“with” (plant species map

from our proposed method)

Green foxtail 72.24 88.71 79.63

80.32

Apple mint 95.05 93.25 94.14

Green pepper 83.09 56.41 67.20

respectively, while 0 or 1 in case of the segmentation

for single images in the middle three columns. As

shown in Figure 4, when a same plant species keeps

present and visible at the same place over weeks, the

value of the corresponding pixels in that plant species

map is increased, displaying brighter. In addition, we

can confirm that even after a plant species is occluded

by other species, we can locate the occluded species

in the map.

Next, we tested the validity of the plant species

map. Ground truth plant species map which indicates

the existence of each plant species for each pixel are

created from annotation data (ground truth of

semantic segmentation) for each time series image.

Table 2 shows the segmentation accuracies of the 6-

th week’s plant species map (“with”) and semantic

segmentation results which is predicted from only one

image of 6-th week without the plant species map

(“without”). According to Table 2, the plant species

map provides more accurate segmentations than

segmentation without using plant species map. In

particular, the Recall for all three classes (plant

species) are significantly lower in the case of the

"without" results, which using the 6-th week’s image

only, without using plant species map, indicating that

many plants are missed. The plant species map, being

additive, yields a lower Precision due to false

positives. However, the higher F1 score, representing

overall accuracy, suggests its effectiveness in

identifying complex vegetation.

4.5 Overgrown Plants Detection

Figure 5 shows examples of the result of using the

plant species map to detect areas occluded by

dominant species. In Figure 5, apple mint and green

foxtail occlude green pepper. From a single-week

segmentation result, locating occluded plant species

under overgrown conditions is challenging. However,

by performing the algorithm, explained in Section 3.3,

for the plant species map and segmentation result,

even plants that are completely occluded can be

detected.

Thus, we have confirmed that the plant species

map provides useful information for pruning. Our

future step is to develop a specific pruning algorithm

using the information obtained from the plant species

map.

Detecting Overgrown Plant Species Occluding Other Species in Complex Vegetation in Agricultural Fields Based on Temporal Changes in

RGB Images and Deep Learning

271

Figure 5: Results of overgrown plants detection.

5 DISSCUSION

To confirm the effectiveness of the proposed method

for detecting overgrown, occluding plant species

(Section 3.3), in 7-th week, we manually pruned the

dominant plants occluding the useful species (green

pepper) so that the useful species is visible. Figure 6

compares the location of the green peppers obtained

by Section 3.3’s method, which uses the plant species

map in the 6-th week, with the green peppers in the

segmentation result of the RGB image after pruning

apple mint and green foxtail in 7-th week as

mentioned earlier. Figure 6 shows that the position of

the green pepper obtained by the plant species map is

almost same as the actual position of the green

pepper. In particular, the locations of the roots of the

green pepper have not changed since the time of

sprouting; so, it can be said that maintaining the past

positional information by time series information is

effective for understanding vegetation in an

environment with severe occlusion.

Figure 6: Checking the position of the green peppers.

6 CONCLUSION AND FUTURE

WORK

This paper has proposed a method for detecting

overgrown plant species occluding other species from

bird's eye view camera images acquired in

Synecocluture farms. Conventional methods for

identifying vegetation from a single image using

semantic segmentation based on deep learning has

difficulty in identifying complex vegetations in

Synecoculture environments. To tackle this issue, our

proposed method consists of acquiring time series

RGB images, performing semantic segmentation for

each of the time series images, creating plant species

map, and detecting overgrown species occluding

other species.

The bird’s eye view camera acquires time series

images at a constant interval (in this paper, once a

week) soon after seeding. We then apply semantic

segmentation to each of the weekly images using a

deep learning model so that areas of the plant species

(in this paper, “apple mint”, “green foxtail” and

“green pepper”) are segmented. The plant species

map consists of multiple layers corresponding to the

segmented species, and at each pixel in each layer, the

number of the existence of that plant species over the

weeks is stored. Finally, we combine the semantic

segmentation results with the (one week) earlier plant

species map so that occluding overgrown species and

occluded species are detected.

Results of experiments using the 36 (= 6 images

(acquired at 6 places) times 6 weeks) time series

images are summarized as follows.

 It turns out that UNet using Resnet101 as the

backbone achieves higher segmentation

accuracies than Resnet50 as the backbone or

Deeplab V3+. UNet-Resnet101 is decided to be

used for the semantic segmentation for each of

time series images.

 As a result of comparing results of semantic

segmentation “with” and “without” the plant

species map, “with” achieves much higher

accuracies than “without”.

 Using the segmentation result and plant species

map, areas of occluding, overgrown species and

occluded species are successfully detected.

In this paper, only green pepper was targeted as a

useful species and apple mint and green foxtail as

dominant species. However, since more plant species

are mixed and densely populated in Synecoculture

environments, our future work includes developing a

method that can deal with a larger number of densely

mixed plant species. In addition, based on this, a

method for determining the positions to be pruned by

our robotic system needs to be achieved.

ACKNOWLEDGEMENTS

This study was conducted with the support of the

Research Institute for Science and Engineering,

Waseda University; Future Robotics Organization,

Waseda University. This work was also supported by

Sustainergy Company (based in Japan), Sony

Computer Science Laboratories, Inc and New Energy

and Industrial Technology Development

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

272

Organization (NEDO). We thank all these

organizations for the financial and technical support

provided. Synecoculture™ is a trademark of Sony

Group Corporation.

REFERENCES

Bendiabdallah, M, H., and Settouti, N. (2021). A

comparison of UNet backbone architectures for the

automatic white blood cells segmentation. WAS

Science Nature (WASSN) ISSN: 2766-7715, 4(1).

Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F., and

Adam, H. (2018). Encoder-decoder with atrous

separable convolution for semantic image

segmentation. Proceedings of the European conference

on computer vision (ECCV):801-818.

Conway, G. R., and Barbie, E. B. (1988). After the green

revolution: sustainable and equitable agricultural

development. Futures, 20(6):651-670.

Geiger, F., Bengtsson, J., Berendse, F., Weisser, W. W.,

Emmerson, M., Morales, M. B., Ceryngier, P., Liira, J.,

Tscharntke, Teja., Winqvist, Camilla., Eggers, S.,

Bommarco, R., Pärt, T., Bretagnolle, V., Plantegenest,

M., Clement, L., Dennis, C., Palmer, C., Oñate, J.,

Guerrero, I., and Inchausti, Pablo. (2010). Persistent

negative effects of pesticides on biodiversity and

biological control potential on European farmland.

Basic and Applied Ecology, 11(2):97-105.

Ide H., Aotake S., Ogata H., Ohtani T., Takanishi A.,

Funabashi M. (2022). A method for detecting dominant

plants in fields from RGB images using Deep Learning.

The Institute of Image Electronics Engineers of Japan:

186-192 (in Japanese).

Kolhar, S., and Jagtap, J. (2023). Phenomics for Komatsuna

plant growth tracking using deep learning approach.

Expert Systems with Applications, 215, 119368.

Li, G., Cui, J., Han, W., Zhang, H., Huang, S., Chen, H.,

and Ao, J. (2022). Crop type mapping using Time series

Sentinel-2 imagery and UNet in early growth periods in

the Hetao irrigation district in China. Computers and

Electronics in Agriculture, 203, 107478.

Mortensen, A. K., Dyrmann, M., Karstoft, H., Jørgensen,

R. N., and Gislum, R. (2016). Semantic segmentation

of mixed crops using deep convolutional neural

network. CIGR-AgEng conference:26-29.

Nasiri, A., Omid, M., Taheri-Garavand, A., and Jafari, A.

(2022). Deep learning-based precision agriculture

through weed recognition in sugar beet fields.

Sustainable Computing. Informatics and Systems, 35,

100759.

Norris, K. (2008). Agriculture and biodiversity

conservation: opportunity knocks. Conservation letters,

1(1):2-11.

Picon, A., San-Emeterio, M. G., Bereciartua-Perez, A.,

Klukas, C., Eggers, T., and Navarra-Mestre, R. (2022).

Deep learning-based segmentation of multiple species

of weeds and corn crop using synthetic and real image

datasets. Computers and Electronics in Agriculture,

194, 106719.

Raja, R., Nguyen, T. T., Slaughter, D. C., and Fennimore,

S. A. (2020). Real-time robotic weed knife control

system for tomato and lettuce based on geometric

appearance of plant labels. Biosystems Engineering,

194:152-164.

Saikawa, T., Cap, Q. H., Kagiwada, S., Uga, H., and

Iyatomi, H. (2019, December). Aop: An anti-overfitting

pretreatment for practical image-based plant diagnosis.

2019 IEEE International Conference on Big Data (Big

Data):5177-5182.

Savci, S. (2012). Investigation of effect of chemical

fertilizers on environment. Apcbee Procedia, 1:287-

292.

Sharifzadeh, S., Tata, J., Sharifzadeh, H., and Tan, B.

(2020). Farm area segmentation in satellite images

using DeeplabV3+ neural networks. Data Management

Technologies and Applications. 8th International

Conference, DATA 2019, Prague, Czech Republic,

July 26–28, 2019, Revised Selected Papers 8:115-135.

Tanaka T., Masaya K., Terae K., Mizukami H., Murakami

M., Yoshida S., Aotake S., Funabashi M., Otani T., and

Takanishi A. (2022). Development of the Agricultural

Robot in SynecocultureTM Environment (1st Report,

Development of Moving Mechanism on the Farm and

Realization of Weeding and Harvesting), Journal of the

Robotics Society of Japan, 40(9):845-848 (in

Japanese).

Tudi, M., Daniel Ruan, H., Wang, L., Lyu, J., Sadler, R.,

Connell, D., Chu, C., and Phung, D. T. (2021).

Agriculture development, pesticide application and its

impact on the environment. International journal of

environmental research and public health, 18(3), 1112.

Wang, C., Du, P., Wu, H., Li, J., Zhao, C., and Zhu, H.

(2021). A cucumber leaf disease severity classification

method based on the fusion of DeeplabV3+ and UNet.

Computers and Electronics in Agriculture, 189,

106373.

Yu, H., Che, M., Yu, H., and Zhang, J. (2022).

Development of Weed Detection Method in Soybean

Fields Utilizing Improved DeeplabV3+ Platform.

Agronomy, 12(11), 2889.

Zou, K., Chen, X., Wang, Y., Zhang, C., and Zhang, F.

(2021). A modified UNet with a specific data

argumentation method for semantic segmentation of

weed images in the field. Computers and Electronics in

Agriculture, 187, 106242.

Detecting Overgrown Plant Species Occluding Other Species in Complex Vegetation in Agricultural Fields Based on Temporal Changes in

RGB Images and Deep Learning

273