A time-series image analysis method and framework for predicting plant growth cycles

Abstract

This invention discloses a time-series image analysis method and framework for predicting plant growth cycles, belonging to the field of image processing technology. It includes: preprocessing collected plant growth image samples to obtain data-augmented samples; constructing a recursive spatiotemporal prediction model based on differential spatiotemporal memory; inputting the data-augmented samples into the recursive spatiotemporal prediction model for training; evaluating the training results based on image quality indicators and visualizing the trained images; deriving the recursive spatiotemporal prediction model as inference weights, and acquiring visualized images in real time for rolling prediction. This invention avoids the damage to key phenotypic features caused by traditional global pooling operations, significantly improving both the structural integrity and visual realism of the predicted images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a time-series image analysis method and framework for predicting plant growth cycles. Background Technology

[0002] Plant growth cycle prediction is one of the core tasks in the field of smart agriculture. It aims to predict the crop morphology in the future by modeling the historical growth image sequence of crops, thereby achieving accurate judgment of crop growth trends and providing data support for refined production management, plant phenotypic analysis and facility cultivation decisions.

[0003] Dicotyledonous crops (such as Arabidopsis thaliana and tomato) are characterized by broad, spreading leaves, significant non-rigid deformation, and a markedly non-stationary growth rate with varying developmental stages. Their growth image sequences contain highly complex evolutionary patterns across time and space. Traditional methods based on manual observation or simple image processing struggle to capture the continuous temporal dynamics of crop growth and cannot be extended to large-scale intelligent applications. In recent years, with the rapid development of computer vision and deep learning technologies, crop growth image prediction based on time-series prediction models has gradually become a research hotspot.

[0004] Existing research on plant growth time-series image prediction can be broadly categorized into four paths: "recursive spatiotemporal memory models," "attention / Transformer-style global dependency modeling," "frequency domain or multi-scale detail enhancement," and "general training frameworks / lack of agriculture-specific evaluation protocols." These methods generally face challenges in long-term prediction of dicotyledonous crops (Arabidopsis thaliana, tomato, etc.): difficulty in characterizing mutations caused by non-stationary growth stages (budding, leaf expansion, organ formation); and the gradual decay of high-frequency details such as leaf margin contours and vein textures during multi-step recursion.

[0005] In addition, existing machine learning frameworks lack dedicated component encapsulation for agricultural time-series tasks, standardized benchmark data interfaces, unified training and testing hyperparameter management mechanisms, and multi-model ensemble evaluation tools. This results in low code reuse rates and high migration costs when researchers conduct tasks such as crop growth prediction, making it difficult to efficiently support comparative experiments on various agricultural time-series tasks and diverse network structures.

[0006] Therefore, a time-series image analysis method and framework for predicting plant growth cycles is needed to solve the above technical problems. Summary of the Invention

[0007] The purpose of this invention is to propose a time-series image analysis method and framework for predicting plant growth cycles.

[0008] A time-series image analysis method for predicting plant growth cycles includes the following steps:

[0009] The collected plant growth image samples were preprocessed to obtain data-enhanced samples;

[0010] Construct a recursive spatiotemporal prediction model based on differential spatiotemporal memory;

[0011] The data-augmented samples are input into the recursive spatiotemporal prediction model for training;

[0012] The training results are evaluated based on image quality metrics, and the trained images are visualized.

[0013] The recursive spatiotemporal prediction model is exported as inference weights, and real-time visual images are collected for rolling prediction.

[0014] Furthermore, the collected plant growth image samples are preprocessed to obtain data-enhanced samples, specifically including:

[0015] The collected plant growth images are sorted chronologically to form the original images. A general segmentation model is used to segment the original images to obtain standardized images containing only the main body of the plant. Using the first frame or reference frame of the standardized images as the target, the images at other times are color-corrected to obtain corrected images. The corrected images are uniformly scaled to a set resolution and organized into five-dimensional tensor samples in chronological order. Sample pairs are constructed using a sliding window method. Data augmentation is performed on the sample pairs by rotation, scaling, translation, and saturation perturbation to obtain data-augmented samples.

[0016] Furthermore, the resolution is set to 192×128 or 192×192.

[0017] Furthermore, the recursive spatiotemporal prediction model based on differential spatiotemporal memory includes a basic spatiotemporal feature extraction layer and several RDMN Cell layers. Each layer sequentially passes through a differential memory unit, a frequency domain global focusing module, a self-attention memory module, and a motion guidance module to complete feature updates, and finally the output is passed to the decoder to generate a predicted image.

[0018] Furthermore, the differential memory unit consists of two sub-modules, MIM-N and MIM-S. MIM-N receives differential features and extracts and enhances non-stationary change components through a gating update process. MIM-S receives the output of MIM-N and integrates and models the stable evolution trend of plant growth in conjunction with the current cell memory state.

[0019] Furthermore, the frequency domain global focusing module first projects the input features to a high-dimensional feature space through a feature mapping layer; then it uses FFT to transform the spatial features to the frequency domain, and achieves global information interaction through frequency domain filtering; finally, it restores the features to the spatial domain through inverse FFT and fuses them with the original features for output.

[0020] A time-series image analysis framework for predicting plant growth cycles includes:

[0021] The data processing layer is used to preprocess the collected plant growth image samples to obtain data-enhanced samples.

[0022] The model building layer is used to build a recursive spatiotemporal prediction model based on differential spatiotemporal memory.

[0023] The training engine layer is used to input data augmentation samples into the recursive spatiotemporal prediction model for training.

[0024] The evaluation visualization layer is used to evaluate the training results based on image quality metrics and visualize the trained images.

[0025] The rolling prediction layer is used to derive inference weights from the recursive spatiotemporal prediction model and perform rolling predictions by acquiring and visualizing images in real time.

[0026] The beneficial effects of this invention are as follows:

[0027] 1. RDMN, by introducing the MIM differential memory mechanism, explicitly models the feature changes between adjacent time steps, accurately capturing non-stationary mutation signals such as new leaf germination in Arabidopsis, effectively addressing the problem of insufficient modeling of non-stationary growth dynamics in existing methods. On the Arabidopsis dataset, RDMN reduces MSE by 64.01% and improves SSIM by 3.89% compared to the baseline model; on the tomato dataset, it reduces MSE by 18.94%, demonstrating significant performance advantages. By introducing an FFT-based SGFM module, it expands the receptive field while preserving high-frequency texture details such as leaf margins and veins, avoiding the damage to key phenotypic features caused by traditional global pooling operations, and significantly improving the structural integrity and visual realism of predicted images.

[0028] 2. The AgriSTL framework achieves high reusability of agricultural time-series analysis code through modular decoupling design, significantly reducing the integration and migration costs of different network structures.

[0029] 3. This invention organically combines a high-precision time-series image prediction method with an efficient agricultural analysis framework, forming a complete technical system from algorithm development to engineering implementation. It provides directly applicable technical support for crop growth status monitoring, growth trend prediction, and refined production management in the field of smart agriculture, and promotes the deep evolution of agricultural artificial intelligence from general methods to agricultural specialization. Attached Figure Description

[0030] Figure 1 This is a flowchart of the time-series image analysis method for predicting plant growth cycles according to the present invention;

[0031] Figure 2A schematic diagram showing the results of extracting the region of interest for plants;

[0032] Figure 3 Comparison chart of color consistency correction effects;

[0033] Figure 4 This is a diagram illustrating data augmentation.

[0034] Figure 5 (a), (b), and (c) are schematic diagrams of three agronomic phenotypic indicators: convex hull area, projected leaf area, and maximum caliper diameter, respectively.

[0035] Figure 6 This is a schematic diagram of the overall structure of RDMN;

[0036] Figure 7 This is a schematic diagram of the MIM Block structure;

[0037] Figure 8 This is a schematic diagram of the SGFM module structure;

[0038] Figure 9 This is a diagram showing the overall structure of the AgriSTL framework.

[0039] Figure 10 This is a schematic diagram of AgriSTL modular integration;

[0040] Figure 11 (a), (b), (c), and (d) are the ablation experimental results of RDMN on the Arabidopsis dataset (pre5_aft5), respectively, for convex hull area, maximum caliper diameter, projected leaf area index, and realism.

[0041] Figure 12 The figure shows the ablation experiment results of RDMN on the Arabidopsis dataset (pre5_aft5);

[0042] Figure 13 (a), (b), (c), and (d) are the comparative experimental results of RDMN on the Arabidopsis dataset (pre5_aft5), representing the convex hull area, maximum caliper diameter, projected leaf area index, and realism, respectively.

[0043] Figure 14 The figure shows the ablation experiment results of RDMN on the tomato dataset (pre5_aft5);

[0044] Figure 15 The figure shows the comparative experimental results of RDMN on the Arabidopsis dataset (pre5_aft5). Detailed Implementation

[0045] This invention proposes a time-series image analysis method and framework for predicting plant growth cycles. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 The flowchart of the time-series image analysis method for predicting plant growth cycles according to the present invention is as follows:

[0047] First, the collected plant growth image samples are preprocessed to obtain data-enhanced samples.

[0048] The plant growth time-series images involved in this invention can be derived from self-collected sequences under protected cultivation conditions, or they can be used in conjunction with publicly available phenotypic data (such as Deep Phenotyping Arabidopsis multi-line sequences). Data acquisition and sample construction follow a streamlined process of "imaging specifications—plant segmentation—light normalization—temporal organization—enhancement and segmentation" to ensure comparability and reproducibility between different time steps and different plants.

[0049] (1) Imaging and Acquisition Methods: In greenhouse or plant factory environments, fixed-position overhead imaging is used to continuously observe individual crops. Taking tomato seedlings as an example, growth images of the same plant can be acquired at fixed time intervals (e.g., once every 12 hours, for about 20 consecutive days) to form an image sequence ordered by time. Public data such as Arabidopsis thaliana usually have multi-time point shooting records under a fixed diurnal rhythm. During the acquisition process, camera parameters, shooting distance, and supplemental lighting conditions should be controlled to reduce apparent changes caused by non-growth factors.

[0050] Plant region extraction: The original image often contains background elements such as the substrate and pot wall. A general segmentation model such as Segment Anything Model (SAM) can be used to generate a plant mask, retain the region of interest, and uniformly black out the background to obtain a standardized image containing only the main body of the plant, thus reducing the interference of background noise on temporal modeling.

[0051] (2) Figure 2 This is a schematic diagram of the results of extracting the region of interest for plants, with the original image at the top and the segmentation and background normalization results at the bottom.

[0052] Color consistency correction: Color temperature differences may exist between early and late periods within the same sequence. A statistical matching / color transfer method based on the LAB color space can be used to perform color correction on images at other times, using the first frame or reference frame of the sequence as the target, to prevent the model from mislearning changes in illumination as changes in growth.

[0053] (3) Figure 3 This is a comparison image showing the effect of color consistency correction, with the top image showing before correction and the bottom image showing after correction.

[0054] (4) Construction and partitioning of time-series samples: The preprocessed images are uniformly scaled to a set resolution (e.g., 192×128 or 192×192) and organized into five-dimensional tensor samples (batch size B, time length T, channels C, height H, width W) in chronological order. A sliding window method is used to construct "historical input frame - future prediction frame" sample pairs, such as pre5_aft5 (5 input frames, 5 prediction frames) or pre4_aft2~6, etc. The dataset should be partitioned by plant as a unit for train / val / test (e.g., 70% / 15% / 15%) to avoid the same plant appearing in both the training and test sets at the same time, which would cause information leakage. Table 1 is a statistical table of experimental dataset construction and configuration.

[0055] Table 1. Statistics on the Construction and Configuration of Experimental Datasets

[0056]

[0057] (5) Data augmentation: During the training phase, image sequences can be augmented with rotation, scaling, translation, saturation perturbation, etc., to improve the robustness of the model to shooting angle, slight displacement and illumination fluctuations; the types, parameter ranges and design purposes of the augmentation strategies are shown in the table below. Table 2 shows the data augmentation strategy design.

[0058] Table 2 Data Augmentation Strategy Design

[0059]

[0060] Figure 4 This is a diagram illustrating data augmentation.

[0061] (6) Phenotypic indicators used for evaluation: In addition to image quality indicators such as MSE, MAE, SSIM, and PSNR, agronomic phenotypic indicators such as projected leaf area, convex hull area, maximum caliper diameter, and solidity can be calculated from the predicted and real sequences to assess whether the prediction conforms to the growth trend from a biological perspective. Figure 5 (a), (b), and (c) are schematic diagrams of three agronomic phenotypic indicators: convex hull area, projected leaf area, and maximum caliper diameter, respectively.

[0062] II. Constructing a recursive spatiotemporal prediction model based on differential spatiotemporal memory.

[0063] RDMN (Recursive Differential Memory Network) is a differential spatiotemporal memory prediction method for non-stationary growth. It takes a sequence of historical plant growth images as input, models the spatiotemporal evolution law step by step through a hierarchical recursive architecture, and outputs a predicted image of crop morphology at future moments.

[0064] 1. Overall Architecture

[0065] RDMN adopts a hierarchical recursive structure: the first layer is the basic spatiotemporal feature extraction layer, which is responsible for the initial spatiotemporal representation learning; the second layer and above are RDMN Cells. Each layer sequentially goes through the differential memory unit (MIM Block), frequency domain global focusing module (SGFM), self-attention memory module (SAM), and motion guidance module (MGM) to complete feature updates, and finally the output is passed to the decoder to generate the predicted image. Figure 6 This is a schematic diagram of the overall structure of RDMN.

[0066] 2. Differential Memory Block (MIM Block)

[0067] Figure 7 This is a schematic diagram of the MIM Block structure; the MIM Block is used to explicitly model the feature changes between adjacent time steps. In the l-th layer (l≥2), the difference feature diff_h = h(l-1,t) - h(l-1,t-1) between the current time and the hidden state of the previous time step is first calculated. This difference feature directly reflects the changes in the plant growth rate and local mutation information.

[0068] The MIM Block consists of two sub-modules: MIM-N and MIM-S. MIM-N receives differential features and extracts and enhances non-stationary change components (mutations and fluctuations) through a gated update process. MIM-S further receives the output of MIM-N and, in conjunction with the current cell memory state, integrates and models the stationary evolutionary trend of plant growth. The two sub-modules work collaboratively, enhancing the response to local changes while preserving overall temporal consistency, effectively solving the problem of modeling non-stationary mutation signals such as new leaf germination.

[0069] 3. Frequency Domain Global Focusing Module (SGFM)

[0070] SGFM (Spectral Global Focus Module) is designed based on Fast Fourier Transform (FFT) to preserve key edge structure information while enhancing global features. Specifically, SGFM first performs spatial preprocessing on the input features, projecting them into a high-dimensional feature space through a feature mapping layer. Then, it uses FFT to transform the spatial features to the frequency domain, achieving global information interaction through frequency domain filtering (dot multiplication in the frequency domain is equivalent to global circular convolution in the spatial domain, without kernel size limitations). Finally, it restores the features to the spatial domain using inverse FFT and fuses them with the original features for output. SGFM can clearly separate low-frequency global trend features from high-frequency texture details while expanding the receptive field, reducing the damage to high-frequency information such as leaf edges and veins caused by traditional global pooling operations. Figure 8 This is a schematic diagram of the SGFM module structure.

[0071] 4. Self-Attention Memory Module (SAM) and Motor Guidance Module (MGM)

[0072] SAM maps current features to historical spatiotemporal memory as queries, keys, and values, aggregates historical information and updates the spatiotemporal memory state through attention weights, and enhances the ability to model long-term temporal contexts. MGM decomposes motion information into transient changes and trend momentum, and dynamically adjusts the output features to better adapt to non-rigid deformation processes.

[0073] Third, input the data-augmented samples into the recursive spatiotemporal prediction model for training.

[0074] 5. Training Strategies

[0075] Table 3 Hyperparameter Settings

[0076]

[0077] During the experiment, typical configurations of training hyperparameters are shown in Table 3. The optimizer was Adam, and the learning rate was 3×10⁻. 4 In conjunction with cosine annealing scheduling, the batch size is 4, and the number of training rounds is set according to crop type; the input resolution and prediction step size (pre / aft) should be consistent with the sequence configuration in Table 1 to ensure that the experiment is reproducible.

[0078] The L2 loss function is used to constrain the pixel-wise error between the predicted sequence and the real sequence. It supports two experimental paradigms: equal-length prediction (such as pre5_aft5) and variable-length prediction (such as pre4_aft2~6). The evaluation metrics include image quality metrics (MSE, MAE, SSIM, PSNR) and agronomic phenotypic metrics (projected leaf area, convex hull area, solidity, etc.).

[0079] 6. Optional phenotypic consistency constraints and multi-task extensions: Phenotypic-related auxiliary losses (such as applying smoothing or consistency constraints to the projected leaf area sequence) can be introduced in addition to pixel-level L2 loss to enhance the consistency between prediction results and agronomic indicators; exogenous sequences such as weather and environmental factors can also be used as conditional inputs to achieve conditional prediction or multi-task joint training.

[0080] Fourth, evaluate the training results based on image quality metrics and visualize the trained images.

[0081] It employs image quality metrics such as SSIM and PSNR, and also supports the automatic calculation and report generation of agronomic phenotypic metrics (projected leaf area, convex hull area, maximum caliper diameter, solidity, etc.). The visualization module supports comparative display of predicted sequences, generation of error heatmaps, and plotting of agronomic phenotypic curves.

[0082] 5. Export the recursive spatiotemporal prediction model as inference weights, and collect and visualize images in real time for rolling prediction.

[0083] After training, the model can be exported as inference weights, which can be used to perform rolling predictions on real-time acquired image sequences at edge computing nodes or servers. For long sequence scenes, sliding window and overlapping stitching strategies can be used to reduce boundary effects, and uncertainty estimation (such as multiple dropout sampling or ensemble models) can be combined to provide risk warnings for production decisions.

[0084] This embodiment also provides a time-series image analysis framework for predicting plant growth cycles, including:

[0085] The data processing layer is used to preprocess the collected plant growth image samples to obtain data-enhanced samples.

[0086] The model building layer is used to build a recursive spatiotemporal prediction model based on differential spatiotemporal memory.

[0087] The training engine layer is used to input data augmentation samples into the recursive spatiotemporal prediction model for training.

[0088] The evaluation visualization layer is used to evaluate the training results based on image quality metrics and visualize the trained images.

[0089] The rolling prediction layer is used to derive inference weights from the recursive spatiotemporal prediction model and perform rolling predictions by acquiring and visualizing images in real time.

[0090] The following is a detailed design of this framework:

[0091] AgriSTL is a machine learning analysis framework for agricultural time-series tasks. AgriSTL (Agricultural Spatiotemporal Learning) is a modular machine learning framework for time-series data analysis in the agricultural field. It provides a complete development loop from data preprocessing, model integration, training and testing to result evaluation, effectively reducing the research and application threshold for agricultural time-series tasks.

[0092] 1. Overall Framework Architecture

[0093] AgriSTL employs a modular and decoupled design, divided into five core layers: data access layer, model component layer, training engine layer, evaluation and visualization layer, and task configuration layer. These layers are connected via standardized interfaces, supporting the free combination and replacement of components.

[0094] Figure 9 This is a diagram showing the overall structure of the AgriSTL framework. Figure 10 This is a schematic diagram of AgriSTL modular integration.

[0095] 2. Data Processing Layer

[0096] The framework provides a standardized interface for accessing agricultural time-series image data, supporting unified format access for both self-collected datasets and public datasets (such as the Deep Phenotyping Arabidopsis dataset). It incorporates a sliding window sample construction strategy, supporting plant-based training / validation / test set partitioning (70% / 15% / 15%), and various prediction length configurations including constant length (pre5_aft5) and variable length (pre4_aft2~6). The data preprocessing module supports image normalization, mask segmentation, color correction, and integrates data augmentation strategies (random rotation, translation, saturation changes, etc.).

[0097] 3. Model Building Layer

[0098] The model building layer is modularly encapsulated using neural network layers as the smallest unit. Currently, 12 mainstream time series prediction network architectures have been integrated and validated, including: ConvLSTM, PredRNNv2, SwinLSTM, TAU, SimVP series, PhyDet, Timesnet, Timemixer, iTransformer, Predformer, VMRNN, etc., and this number is continuously expanding to at least 15. Each network architecture is encapsulated as a component with a unified interface, supporting free invocation via configuration files. The RDMN model proposed in this invention has been integrated as an important component of the framework.

[0099] 4. Training the engine layer

[0100] The training engine layer provides unified management of training, validation, and testing processes. It supports configurable management of hyperparameters (learning rate, batch size, optimizer, scheduling strategy, etc.), multi-GPU parallel training, and provides mechanisms for resuming training after breakpoints and saving model checkpoints. The unified training interface shields researchers from the differences in implementations of different models, allowing them to focus solely on model design and parameter configuration.

[0101] 5. Evaluate the visual layer

[0102] The framework incorporates a dedicated evaluation protocol for agricultural time-series tasks, prioritizing image quality metrics such as SSIM and PSNR. It also supports the automatic calculation and report generation of agronomic phenotypic metrics (projected leaf area, convex hull area, maximum caliper diameter, and solidity). The visualization module supports comparative display of predicted sequences, generation of error heatmaps, and plotting of agronomic phenotypic curves, facilitating researchers' intuitive evaluation of model performance.

[0103] 6. Rolling Prediction Layer

[0104] The framework currently stably supports two typical agricultural time-series tasks (plant growth image prediction and weather time-series prediction), and plans to expand to at least three. By specifying the task type, data path, model structure, and evaluation protocol through task configuration files, cross-task migration can be achieved with "zero code modification," significantly reducing research migration costs.

[0105] The RDMN method of this invention explicitly models the feature changes between adjacent time steps by introducing the MIM differential memory mechanism, accurately capturing non-stationary mutation signals such as the germination of new leaves in Arabidopsis thaliana, effectively solving the problem of insufficient modeling of non-stationary growth dynamics in existing methods. On the Arabidopsis thaliana dataset, RDMN reduces MSE by 64.01% and improves SSIM by 3.89% compared to the baseline model; on the tomato dataset, it reduces MSE by 18.94%, demonstrating significant performance advantages. By introducing an FFT-based SGFM module, the receptive field is expanded while preserving high-frequency texture details such as leaf margins and veins, avoiding the damage to key phenotypic features caused by traditional global pooling operations, and significantly improving the structural integrity and visual realism of the predicted images.

[0106] To verify the independent contribution and synergistic effect of the core modules of this invention, ablation experiments were conducted on Arabidopsis thaliana and tomato datasets: four configurations were compared, namely the baseline model, SGFM only, MIM Block only, and complete RDMN (dual module fusion), and the results were verified by image quality indicators, agronomic phenotypic indicators, and visualization results.

[0107] Table 4. Ablation experimental results of RDMN on the Arabidopsis thaliana dataset (pre5_aft5) (image quality metrics)

[0108]

[0109] Table notes: "↓" indicates that the lower the indicator, the better; "↑" indicates that the higher the indicator, the better; bold indicates the best; underline indicates the second best.

[0110] As shown in Table 4, compared to the baseline GMG, after introducing SGFM alone, the MSE of GMG+SGFM decreased from 148.084 to 144.305, the SSIM increased from 0.874 to 0.883, and the PSNR increased from 27.626 to 27.713, while the MAE increased slightly. This indicates that SGFM enhances the structural representation ability of features through frequency domain global information interaction and improves the preservation of high-frequency details to some extent, but its optimization of the overall temporal error is still limited when used alone. In contrast, after introducing the MIM Block, the MSE, MAE, and PSNR of GMG+MIM Block decreased by 63.38%, 37.45%, and 14.26% respectively compared to the baseline, and the SSIM increased to 0.909. This shows that the differential memory mechanism can effectively characterize the higher-order temporal nonstationarity in the Arabidopsis growth process and has a stronger modeling ability for local growth rate changes and continuous evolution trends, which is the main factor improving prediction performance. After further integrating SGFM, RDMN achieved the best results in MAE, MSE, and PSNR. SSIM and GMG+MIM Block were basically on par, indicating that MIMBlock and SGFM have good complementarity in terms of function. The former enhances the ability to model non-stationary time series dynamics, while the latter further improves the global structure representation and high-frequency detail preservation, thereby improving the overall prediction performance of the model.

[0111] Figure 11 (a), (b), (c), and (d) show the ablation experiment results of RDMN on the Arabidopsis dataset (pre5_aft5) for convex hull area, maximum caliper diameter, projected leaf area, and solidity, respectively. The phenotypic curves show that the RDMN model's curves for convex hull area, maximum caliper diameter, and projected leaf area are significantly better than the baseline model and maintain high consistency with GT. Comparing the results of individual modules, the curve of the GMG+MIM Block is generally closer to GT than that of GMG+SGFM, indicating that the MIM differential memory mechanism is the core factor improving prediction performance on the Arabidopsis dataset, and its modeling effect on non-stationary temporal changes is more significant. The solidity index shows that after further combining SGFM with the GMG+MIM Block, the solidity curve of RDMN is closer to the decreasing trend of GT, indicating that the SGFM module helps enhance the model's ability to express plant outline integrity and overall structural morphology.

[0112] Figure 12The figure shows the ablation experiment results of RDMN on the Arabidopsis dataset (pre5_aft5). The figure clearly shows that as the prediction step size increases, the leaf margin boundaries in the baseline GMG model gradually become blurred, and the predicted leaf size is generally smaller than the GT, indicating insufficient characterization of leaf growth expansion and limited ability to model subsequent plant growth trends. In contrast, the GMG+MIM Block model is closer to the GT in overall morphology, indicating that the introduction of the MIM Block differential memory mechanism enhances the model's ability to represent dynamic changes during the non-stationary growth of dicotyledonous crops. When SGFM is introduced alone, GMG+SGFM outperforms the baseline in recovering complex regions in the plant center; RDMN combines the advantages of both, achieving optimal results in both morphological integrity and detail restoration.

[0113] Figure 13 Figures (a), (b), (c), and (d) show the comparative experimental results of RDMN on the Arabidopsis dataset (pre5_aft5) for convex hull area, maximum caliper diameter, projected leaf area, and solidity, respectively. From the agronomic phenotypic curves, the results on the tomato dataset are more complex than those on Arabidopsis. The numerical values ​​of the curves alone are insufficient to fully determine the superiority or inferiority of the models; therefore, a combined analysis with visualization results is necessary. Specifically, for the two more global-scale indicators, convex hull area and maximum caliper diameter, the curve of GMG+SGFM is closer to the ground truth (GT) than the baseline, indicating that SGFM has a positive effect on the overall canopy scale of the plant. However, for local indicators such as projected leaf area, GMG+MIM Block performs better, indicating that the differential memory mechanism is more sensitive to the local growth dynamics of tomato leaves. RDMN shows better synergistic effects on multiple comprehensive phenotypic indicators.

[0114] Table 5. Ablation experiment results of RDMN on the tomato dataset (pre5_aft5) (image quality indicators)

[0115]

[0116] Table notes: "↓" indicates that the lower the indicator, the better; "↑" indicates that the higher the indicator, the better; bold indicates the best; underline indicates the second best.

[0117] As shown in Table 5, on the tomato dataset, introducing SGFM alone improved MSE and PSNR by 14.83% and 2.75% respectively compared to the baseline, indicating that the frequency domain global focusing module helps maintain the overall plant outline and spatial scale. Introducing MIMBlock alone increased SSIM to 0.890, indicating that the differential memory mechanism has a constraining effect on complex non-stationary dynamics and local structural consistency. The RDMN fusion of the two modules achieved the best results in MAE, MSE, and PSNR, demonstrating the synergistic gain of MIM Block and SGFM, and verifying the effectiveness and robustness of the present invention in complex tomato growth scenarios.

[0118] Figure 14 The image shows the ablation experiment results of RDMN on the tomato dataset (pre5_aft5). Visualizing the results helps to explain the above phenomena more clearly. The baseline GMG shows more pronounced leaf edge blurring and size deviation at longer prediction steps such as t+3 and t+5, indicating its limited ability to depict fine-grained morphological evolution during continuous tomato growth. GMG+MIMBlock improves leaf pose changes and local structure preservation, while GMG+SGFM is superior in restoring complex textured areas in the plant center. After fusing the two modules, RDMN's predicted images are closer to GT in terms of leaf edge sharpness, leaf position, and canopy structure.

[0119] Furthermore, the RDMN of this invention was compared with mainstream spatiotemporal prediction models such as TAU, MIM, PredRNNv2, GMG, PhyDNet, MMVP, PredFormer, and the improved baseline DSAVFN. The adaptability and superiority of this invention for plant growth and development image prediction tasks were verified from three dimensions: image quality indicators, agronomic phenotypic indicators, and visualization results.

[0120] Table 6 Comparative experiments of RDMN on the Arabidopsis dataset (pre5_aft5) (image quality metrics)

[0121]

[0122] Table notes: "↓" indicates that the lower the indicator, the better; "↑" indicates that the higher the indicator, the better; bold indicates the best; underline indicates the second best.

[0123] As shown in Table 6, the RDMN of this invention achieves the best results in all four image quality metrics (MAE, MSE, SSIM, and PSNR) on the Arabidopsis dataset. Compared to DSAVFN, MAE is reduced by 37.20%, MSE by 65.19%, SSIM is improved by 2.60%, and PSNR is improved by 14.88%. This indicates that while maintaining the consistency of leaf contour and spatial structure, this invention can more accurately depict the non-stationary growth process of the plant in subsequent time steps, demonstrating significant technical effects and practical value.

[0124] From the agronomic phenotypic index curves, RDMN also showed the best overall performance in terms of bulge area, maximum caliper diameter, projected leaf area, and seed density. The first three indicators reflect the overall extensional changes brought about by plant canopy expansion and leaf growth. RDMN's prediction curves were generally closer to GT at each time point, indicating that it more stably depicted the continuous unfolding of Arabidopsis leaves and the trend of canopy expansion. Regarding seed density, RDMN's seed density curve showed a consistent downward trend with GT for the first time, and the numerical deviation was significantly reduced, indicating that the model can more accurately reflect the actual growth state of the plant canopy gradually shifting from compact to loose.

[0125] Figure 15 The figure shows the comparative experimental results of RDMN on the Arabidopsis dataset (pre5_aft5). As can be seen from the figure, compared with general spatiotemporal prediction models such as TAU, MIM, and PredRNNv2, the RDMN of this invention has advantages in terms of predicted image sharpness, leaf boundary preservation, and new leaf region recovery. Compared with the improved baseline DSAVFN, RDMN can more accurately characterize the appearance location and growth trend of new leaves in subsequent time steps, further demonstrating the modeling ability of differential spatiotemporal memory mechanism for non-stationary growth dynamics.

[0126] The AgriSTL framework of this invention achieves high reusability of agricultural time-series analysis code through modular decoupling design, significantly reducing the integration and migration costs of different network structures. Standardized benchmark interfaces and unified training processes make multi-model comparison experiments more standardized and reproducible. The framework currently integrates 12 mainstream time-series network structures, stably supporting typical agricultural tasks such as plant growth prediction, and has been selected as one of the first 20 benchmarks for the MADA platform, possessing significant platform application and academic promotion value.

[0127] This invention organically combines a high-precision time-series image prediction method with an efficient agricultural analysis framework, forming a complete technical system from algorithm development to engineering implementation. It provides directly applicable technical support for crop growth status monitoring, growth trend prediction, and refined production management in the field of smart agriculture, and promotes the deep evolution of agricultural artificial intelligence from general methods to agricultural specialization.

[0128] This preferred embodiment provides a computer device that can implement the steps of the time-series image analysis method for plant growth cycle prediction provided in the embodiments of this application. Therefore, it can achieve the beneficial effects of the time-series image analysis method for plant growth cycle prediction provided in the embodiments of this application.

[0129] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. Therefore, embodiments of this application provide a storage medium storing multiple instructions that can be loaded by a processor to execute the steps of any embodiment of the time-series image analysis method for plant growth cycle prediction provided in this application.

[0130] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0131] Since the instructions stored in the storage medium can execute the steps in any of the time-series image analysis methods for plant growth cycle prediction provided in the embodiments of this application, the beneficial effects that any of the time-series image analysis methods for plant growth cycle prediction provided in the embodiments of this application can achieve can be realized.

[0132] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A time-series image analysis method for predicting plant growth cycles, characterized in that, Includes the following steps: The collected plant growth image samples were preprocessed to obtain data-enhanced samples; Construct a recursive spatiotemporal prediction model based on differential spatiotemporal memory; The data-augmented samples are input into the recursive spatiotemporal prediction model for training; The training results are evaluated based on image quality metrics, and the trained images are visualized. The recursive spatiotemporal prediction model is exported as inference weights, and real-time visual images are collected for rolling prediction.

2. The time-series image analysis method for predicting plant growth cycles according to claim 1, characterized in that, The collected plant growth image samples are preprocessed to obtain data-enhanced samples, specifically including: The collected plant growth images are sorted chronologically to form the original images. A general segmentation model is used to segment the original images to obtain standardized images containing only the main body of the plant. Using the first frame or reference frame of the standardized images as the target, the images at other times are color-corrected to obtain corrected images. The corrected images are uniformly scaled to a set resolution and organized into five-dimensional tensor samples in chronological order. Sample pairs are constructed using a sliding window method. Data augmentation is performed on the sample pairs by rotation, scaling, translation, and saturation perturbation to obtain data-augmented samples.

3. The time-series image analysis method for predicting plant growth cycles according to claim 2, characterized in that, The set resolution is 192×128 or 192×192.

4. The time-series image analysis method for predicting plant growth cycles according to claim 3, characterized in that, The recursive spatiotemporal prediction model based on differential spatiotemporal memory includes a basic spatiotemporal feature extraction layer and several RDMN Cell layers. Each layer sequentially goes through a differential memory unit, a frequency domain global focusing module, a self-attention memory module, and a motion guidance module to complete feature updates. Finally, the output is passed to the decoder to generate a predicted image.

5. The time-series image analysis method for predicting plant growth cycles according to claim 4, characterized in that, The differential memory unit consists of two sub-modules, MIM-N and MIM-S. MIM-N receives differential features and extracts and enhances non-stationary change components through a gating update process. MIM-S receives the output of MIM-N and integrates and models the stable evolution trend of plant growth in conjunction with the current cell memory state.

6. The time-series image analysis method for predicting plant growth cycles according to claim 4, characterized in that, The frequency domain global focusing module first projects the input features to a high-dimensional feature space through a feature mapping layer; then it uses FFT to transform the spatial features to the frequency domain, and achieves global information interaction through frequency domain filtering; finally, it restores the features to the spatial domain through inverse FFT and fuses them with the original features for output.

7. A time-series image analysis framework for predicting plant growth cycles, characterized in that, The time-series image analysis method for predicting plant growth cycles as described in any one of claims 1 to 6 includes: The data processing layer is used to preprocess the collected plant growth image samples to obtain data-enhanced samples. The model building layer is used to build a recursive spatiotemporal prediction model based on differential spatiotemporal memory. The training engine layer is used to input data augmentation samples into the recursive spatiotemporal prediction model for training. The evaluation visualization layer is used to evaluate the training results based on image quality metrics and visualize the trained images. The rolling prediction layer is used to derive inference weights from the recursive spatiotemporal prediction model and perform rolling predictions by acquiring and visualizing images in real time.

Images