Vegetation index time series prediction system based on ConvLSTM spatiotemporal coding

Abstract

The application discloses a vegetation index time series prediction system based on ConvLSTM space-time coding, comprising: a data preparation module, which acquires historical NDVI time series and performs downsampling and time series space double interpolation processing thereon; a remote sensing image inference prediction module, which comprises: an encoder, a decoder, a jump connection module and a double-channel output head; the encoder is a three-layer stacked ConvLSTM encoder, the decoder is a three-layer ConvLSTM autoregressive decoder structure, a channel attention and spatial attention module is embedded on the decoder output feature path, the decoder fuses the multi-scale spatial detail information of the encoder through jump connection, and the double-channel output head outputs the NDVI time series prediction value of each point and the uncertainty logarithmic representation of each prediction point. The application has higher prediction accuracy, possesses pixel-by-pixel uncertainty quantification capability, missing data robust processing and large-area seamless inference capability.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation of remote sensing images and deep learning, specifically to a method and system for time-series prediction of vegetation indices. Background Technology

[0002] Satellite remote sensing, with its large-scale, periodic, and low-cost Earth observation capabilities, has become an irreplaceable data source for monitoring vegetation dynamics at regional and even global scales. The Normalized Difference Vegetation Index (NDVI), as a core indicator reflecting the intensity and growth status of vegetation photosynthesis, provides high temporal resolution and long-term time-series predictions. This is a key technological foundation for applications such as monitoring crop growth and yield forecasting, early warning of drought and pests, assessment of reforestation and grassland restoration, and vegetation carbon sink accounting under carbon neutrality objectives.

[0003] Existing methods for time-series vegetation index prediction are mainly divided into two categories: physical process models and data-driven models. Physical process models predict vegetation growth by simulating the interaction mechanism between vegetation growth and environmental factors, offering strong physical interpretability. However, they typically require a large number of difficult-to-obtain physiological parameters and meteorological inputs, and their spatial generalization ability is limited. Data-driven models, on the other hand, learn temporal evolution patterns directly from historical observations, without the need for complex prior parameterization. However, traditional statistical regression and shallow machine learning methods are limited by their ability to model long-term nonlinear dynamics, making it difficult to capture the multi-scale, strongly coupled temporal dependencies in vegetation growth. Their prediction accuracy decreases sharply with the increase of the prediction step size.

[0004] In recent years, deep learning has demonstrated significant potential in the field of remote sensing time series analysis. Convolutional Neural Networks (CNNs) can effectively extract local spatiotemporal patterns, but are limited by their inherent receptive field, making it difficult to model long-term dependencies across phenological periods. Recurrent Neural Networks (RNNs) and their Long Short-Term Memory variants (LSTMs) are naturally suitable for sequence modeling, but they are prone to gradient vanishing in long sequences, and recursive serial computation limits training and inference efficiency. Transformers capture global temporal correlations through self-attention mechanisms and have been widely introduced into remote sensing time series tasks in recent years; however, their computational complexity increases quadratically with sequence length, resulting in high computational costs for large-area, long-time-series data. Furthermore, self-attention mechanisms are insufficient in characterizing the fine-grained spatial structure of land cover boundaries, which is detrimental to maintaining the edge accuracy of crop plots. Based on these reasons, this invention selects a Convolutional Hybrid Long Short-Term Memory Network (ConvLSTM) as the core temporal modeling unit. This architecture extracts spatial neighborhood features through convolutional operations and captures temporal dependencies through LSTM state propagation, maintaining linear computational complexity while ensuring spatial detail fidelity, making it more suitable for the operational needs of large-area vegetation index prediction.

[0005] More importantly, existing deep learning time series prediction methods still face three core bottlenecks that have not been fully addressed in practical agricultural and ecological applications. First, in the iterative decoding process of mainstream autoregressive prediction strategies, small single-step prediction errors accumulate and amplify, leading to significant deviations from true values ​​in later long-term predictions and rendering long-term predictions worthless. Second, existing models generally only output deterministic point-scale predictions, lacking the ability to assess the confidence level of the prediction results. In high-risk scenarios such as agricultural disaster early warning and ecological management decision-making, predictions lacking confidence assessment are insufficient to support reliable decision-making. Third, when facing operational inference at the provincial or larger scale, the spatiotemporal gaps caused by cloud cover in the input data and the boundary artifacts introduced by tile-based inference intertwine, severely restricting the engineering usability of the prediction results. Summary of the Invention

[0006] To address the problems in the background technology, this invention proposes a vegetation index time series prediction system, which is based on a convolutional long short-term memory network (ConvLSTM) encoder-decoder architecture and integrates the vegetation index (NDVI) with adaptive uncertainty quantization.

[0007] This invention proposes a vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding, comprising: a data preparation module, which acquires historical normalized vegetation index (NDVI) time series and performs downsampling and spatiotemporal dual interpolation processing on the historical NDVI time series; and a remote sensing image inference prediction module, which includes: an encoder, a decoder, a skip connection module, and a dual-channel output head; the encoder is a three-layer stacked ConvLSTM encoder, and the decoder is a three-layer ConvLSTM autoregressive decoder structure. The encoder and decoder are fused with multi-scale spatial details through the skip connection module to alleviate the blurring of deep feature space. Channel attention and spatial attention modules are embedded after the feature output of each layer of the decoder to achieve dual-path spatiotemporal feature enhancement. The dual-channel output head outputs two prediction results: the first channel is the NDVI time-series prediction value for each point, and the second channel is the logarithmic representation of the uncertainty for each prediction point.

[0008] This invention combines a Convolutional Long Short-Term Memory (ConvLSTM) network with a dual-channel uncertain output encoder-decoder prediction network. The overall structure employs a three-layer ConvLSTM encoder and a three-layer ConvLSTM autoregressive decoder. Channel attention and spatial attention dual-path enhancement modules are embedded in the encoder-decoder feature transfer path to capture multi-scale spatiotemporal dependencies. In the decoder stage, skip connections are used to fuse multi-scale spatial details from the encoder to alleviate the problem of deep feature space ambiguity. For uncertainty modeling, a dual-channel output head is designed to output the NDVI prediction value and log-variance respectively. Gaussian negative log-likelihood loss is used for probabilistic prediction training, enabling the model to provide a quantitative confidence estimate while outputting pixel-by-pixel prediction values. Addressing the stitching artifact problem in large-area remote sensing image inference, a two-dimensional Hann window-based overlapping block weighted stitching method is proposed to achieve seamless large-area prediction.

[0009] This invention addresses the problems in existing NDVI time-series prediction methods, such as unreliable long-term predictions due to autoregressive error accumulation, difficulty in supporting high-risk decision-making due to lack of prediction confidence metrics, and constraints on operational applications caused by missing spatiotemporal data and artifacts at the boundaries of large-area block inference. The NDVI time-series prediction method proposed in this invention has higher prediction accuracy than currently popular methods, and also possesses pixel-by-pixel uncertainty quantification capabilities, robust handling of missing data, and seamless inference capabilities over large areas, providing reliable confidence bases for risk-sensitive scenarios such as precision agriculture and ecological monitoring. Attached Figure Description

[0010] To facilitate understanding of the invention, it will be described in more detail with reference to the specific embodiments shown in the accompanying drawings. These drawings depict only typical embodiments of the invention and should not be considered as limiting the scope of protection of the invention.

[0011] Figure 1 This is a flowchart of the method of the present invention.

[0012] Figure 2 This is a flowchart illustrating the overall architecture of the system of the present invention.

[0013] Figure 3 A graph comparing the predicted NDVI value, the actual value, and the time of uncertainty for a single instance.

[0014] Figure 4 A graph comparing the predicted NDVI value, the actual value, and the uncertainty space for a single instance. Detailed Implementation

[0015] The embodiments of the present invention are described below with reference to the accompanying drawings to enable those skilled in the art to better understand and implement the present invention. However, the listed embodiments are not intended to limit the present invention. In the absence of conflict, the following embodiments and the technical features in the embodiments can be combined with each other, wherein the same components are indicated by the same reference numerals.

[0016] The following reference Figures 1-2 This invention is described.

[0017] Example 1 The method of the present invention includes S1: performing step-size sampling and data preprocessing on historical NDVI time series.

[0018] Historical NDVI time series (T in The process of acquiring ×H×W) data is as follows: acquire NDVI time series remote sensing images of the target area. The data comes from satellite sensors such as MODIS, Landsat, or Sentinel-2 and is stored in GeoTIFF format with dimensions of (total days, height, width). Each pixel value is a floating-point number in the range of [-1, 1]. Invalid observations (cloud obstruction, atmospheric interference) are marked with a specified NoData value (such as -9999.0 or -1.0) or directly stored as NaN (Not a Number, which is an outlier).

[0019] S1.1 employs a dual-stride sampling strategy, downsampling the input and output sequences with a step size greater than 1.

[0020] Traditional methods input all frames of a continuous time series into the model frame by frame, resulting in: (1) highly redundant information between adjacent time steps, increasing computational overhead; (2) excessively long autoregressive links for output prediction, with each step of autoregressive inference introducing new errors, which accumulate exponentially along the link length.

[0021] This invention employs a completely new method. It is described below.

[0022] Input Stride S in : For the historical sequence every S in Take one frame at a time to form a frame of length T. in The input sequence actually spans a time window of (T) in -1) ×S in +1 day.

[0023] Output Stride S out The number of frames directly predicted by the model is T. pred The inter-frame interval is S outThe actual predicted span is T days. pred ×S out sky.

[0024] The output step size will be the total number of steps in the autoregressive inference, calculated from the actual number of predicted days N. days Compress to N days / S out Step. Let the output step size be S. out =2, the number of autoregressive steps for predicting 20 days is reduced from 20 steps to 10 steps, and the error accumulation coefficient is reduced from approximately O(20) to O(10). Intermediate day sequence frame P t+s Linear interpolation is used for completion, and the specific interpolation method is as follows: , Where s represents the intermediate interpolation step number, S out P represents the interpolation step size of the output sequence. t and P t+1 P represents the prediction results corresponding to two adjacent known time steps. t+s Indicates in P t and P t+1 The s-th intermediate prediction result is obtained through linear interpolation.

[0025] A preamble frame P0 is inserted before the first frame of the output sequence, and the output sequence is formed by the last frame X of the input sequence. last The interpolation obtained from the predicted first frame P1 is as follows: , S1.2 performs temporal-spatial double missing value interpolation on the downsampled input sequence obtained in S1.1.

[0026] In this invention, the input sequence undergoes temporal-spatial dual missing value interpolation. For the output sequence, only intermediate frame interpolation is performed; no further processing is required.

[0027] Cloud occlusion in remote sensing data leads to a large number of NaN values. This invention employs a two-stage interpolation strategy: The first stage (temporal linear interpolation): For each spatial location's temporal vector, linear interpolation is performed along the time axis, filling missing values ​​using valid observations preceding and following the same location. The second stage (spatial nearest neighbor interpolation): For NaN values ​​remaining from the first stage (in cases where the entire column is NaN), a distance transform is used to find the nearest valid pixel for spatial filling. The two stages are complementary: temporal interpolation handles some missing values, while spatial interpolation handles all missing values, ensuring that no NaN values ​​from the input sequence enter the model.

[0028] Preferably, step S1 further includes S1.3 and S1.4.

[0029] S1.3, Normalization and Mask Generation. The NDVI values ​​of the differenced input sequence are cropped to the range of [-1, 1] to generate a valid pixel mask (valid = 1, invalid = 0). NaN is replaced with -1.0 as the model input.

[0030] S1.4, Data Augmentation. The input sequence is randomly augmented with a 50% probability. The main augmentation operations include: horizontal flipping (mirroring along the width dimension), vertical flipping (mirroring along the height dimension), and brightness adjustment (multiplying by a random factor in the range of [0.9, 1.1] and then cropping to [-1, 1]).

[0031] like Figure 1 As shown, the method of the present invention further includes step S2: The historical NDVI time-series data obtained in S1 is processed by a three-layer ConvLSTM encoder with time-step gated convolution, relying on the hidden states to transmit information across time steps, outputting the encoded spatiotemporal feature sequence and the final hidden states of each layer. This spatiotemporal feature sequence and the final hidden states of each layer are used by the subsequent decoder to predict future NDVI.

[0032] Traditional fully connected LSTM flattens the spatial dimension into a one-dimensional vector, which destroys the spatial topological relationship between pixels.

[0033] This invention employs Convolutional LSTM (ConvLSTM), replacing matrix multiplication in the LSTM gating equations with convolution operations, thus preserving the spatial structure of the hidden states. The state update equation for a single-layer ConvLSTM unit is: , in, This represents the convolution operation, and 'e' represents element-wise multiplication. The sigmoid activation function is calculated as follows: X t H represents the feature map input to the ConvLSTM unit at time step t; t-1 H represents the hidden state of the previous time step, used to convey historical time series features; t Indicates the hidden state output at the current time step; C t-1 This represents the cell state at the previous time step and is used to preserve long-term memory information; C t This represents the cell state updated at the current time step. W represents the learnable convolution kernel parameters, including the convolution kernels input to each gate unit and the convolution kernels from the hidden state to each gate unit, including W... xi W hi W xf W hf W xo W ho W xg W hgetc.; b represents the bias term of the corresponding gating unit, including b i b f b o and b g In the formula, all W values ​​are 3×3 convolution kernels. Input gate i t Forgotten Gate t Output gate o t and candidate memory G t Computational efficiency is optimized by performing a joint convolution operation followed by splitting. , In one implementation, a multi-layer stacked encoder is used. Figure 1 The 30 frames of NDVI images obtained from S1 are processed layer by layer through ConvLSTM (displayed as 3 layers). Within each layer, information is transmitted across time by hidden states, and between layers, the output sequences are concatenated to finally output a high-level spatiotemporal feature sequence with 256 channels and the state at the end of each layer.

[0034] Specifically, the encoder of this invention consists of three stacked ConvLSTM layers, with the number of channels increasing sequentially to 64, 128, and 256. Layer 1 (64 channels): captures fine-grained local texture and edge features; Layer 2 (128 channels): models mesoscale vegetation patches and growth patterns; Layer 3 (256 channels): extracts large-scale phenological trends and climate response features. The three ConvLSTM layers are connected in a multi-scale spatiotemporal feature pyramid structure, with the output hidden state sequence of the lower layer serving as the input of the upper layer.

[0035] The method of the present invention also includes S3: starting from the output of the encoder at the last moment, the decoding is performed step by step through 3 layers of ConvLSTM + skip connection + dual path attention, and the predicted value of each step is fed back as the input of the next time step, generating 10 future NDVI frames (including uncertainty variance) through autoregression.

[0036] S3.1, the decoder is used to decode the feature sequence obtained in S2.

[0037] The decoder also consists of three stacked ConvLSTM layers, with the number of channels decreasing sequentially to 256, 128, and 64, forming a symmetrical structure with the encoder. The decoder generates future prediction sequences frame-by-frame in an autoregressive manner. , in This is the predicted value from the previous moment. This is the hidden state for the decoder.

[0038] Decoder initial state Hidden state of the encoder's last layer Generated through a nonlinear predictor: , The decoder's first frame input is the last time step feature output from the encoder's final layer, and subsequent frames' inputs are... The predicted value of the previous frame Projection: , S3.2: The feature sequence obtained in S3.1 is enhanced using a dual-path attention enhancement mechanism: first, channel attention is used to select "which features are important," then spatial attention is used to select "which positions are important," and finally, the enhanced feature map is output for the next layer of decoding or final prediction. This invention introduces a dual-path enhancement module of channel attention and spatial attention on the feature transfer path between the encoder and decoder.

[0039] Channel attention performs global pooling on the feature map, followed by two 1×1 convolutional layers and then sigmoid activation to generate channel weight vectors. This branch adaptively strengthens the feature channels that are important for the prediction task and suppresses noisy channels.

[0040] Spatial attention is calculated along the channel dimension using average pooling and max pooling respectively, and then concatenated into a 2-channel feature map. This map is then passed through a 7×7 convolutional layer followed by Sigmoid activation to generate a spatial attention map. This branch guides the model to focus on areas with drastic vegetation changes (such as farmland and forest edges), reducing attention to uniform areas (such as water bodies and bare soil).

[0041] The final output of attention is a cascade of two types of attention: , in, F W represents the feature map of the input attention module. channel W represents the channel attention weight. spatial Represents spatial attention weights. This indicates element-wise multiplication.

[0042] Preferably, such as Figure 1 As shown, the method of the present invention further includes a skip connection step: fusing multi-scale spatial details between the encoder and the decoder through skip connections to alleviate the blurring of the deep feature space.

[0043] Specifically, the skip connection operation includes: (1) obtaining the feature maps output by each layer of the encoder and caching them as spatial detail features of the corresponding scale; (2) performing channel mapping on the encoder features of each layer using 1×1 convolution to make the number of output channels consistent with the number of output channels of the corresponding decoder layer; (3) fusing the encoder features after channel mapping with the output features of the corresponding decoder layer in an element-wise addition manner to obtain fused features; (4) using the fused features as the output of the corresponding decoder layer and inputting them into the next decoding layer to continue feature recovery.

[0044] In this method, the encoder and decoder establish a skip connection using a mirror correspondence. The encoder's first, second, and third layers are respectively connected to the decoder's third, second, and first layers to achieve the fusion of shallow spatial detail features and deep semantic features.

[0045] In this way, the original spatial details preserved by each layer of the encoder (after aligning the channels through 1×1 convolution) are directly injected into the corresponding layer of the decoder, making up for the fine-grained spatial information lost in the deep features and improving the clarity of the predicted contours.

[0046] The output features of each encoder layer are aligned to the number of channels through a 1×1 convolution and then added to the decoder to form a skip connection: , The decoder is in a hidden state. For encoder output features, Skip connections enable the decoder to directly access high-resolution spatial details in the encoder, effectively mitigating the problem of spatial information becoming blurred as the number of layers increases in multi-layer ConvLSTM.

[0047] like Figure 1 As shown, the method of the present invention further includes S4: dual-channel uncertainty quantization outputs the prediction result.

[0048] Traditional NDVI prediction models output only a single deterministic value (point estimate), failing to differentiate the reliability of the prediction. While the model provides the same output format for clear-sky and cloud-covered areas, the prediction quality differs significantly. This invention upgrades point prediction to probabilistic prediction, making the output for each spatiotemporal location a probability distribution rather than a single scalar.

[0049] This invention uses a dual-channel output head to output dual-channel prediction results: the decoder loops 10 times and outputs the prediction results for the next 10 frames, each frame containing both ndvi prediction values ​​and uncertainties.

[0050] The output convolutional head, located above the last layer of features in the decoder, outputs a dual-channel prediction result. The first channel represents the temporal NDVI prediction value for each point, and the second channel represents the uncertainty logarithm for each prediction point. To ensure numerical stability, the uncertainty logarithm variance channel is pruned.

[0051] , in, This represents the stable representation after taking the logarithm of the prediction variance. This represents the high-level semantic features output by the decoder. The value is -10. The value is 5.

[0052] The output is the predicted NDVI value and the pixel-wise uncertainty (Tout×H×W×2).

[0053] The method of the present invention also includes S5: block weighted splicing inference output.

[0054] S5.1, the overlapping block segmentation strategy is as follows: An overlapping sliding window segmentation strategy is adopted. The block size is set to P × P, the overlap amount to R, and the sliding step length to stride = P - R. For the original image H × W, the required number of blocks is calculated in both the horizontal and vertical directions: , N h N represents the number of blocks required in the vertical direction. w Indicates the number of blocks required in the vertical direction.

[0055] Reflect padding is applied to boundary areas that are not large enough to ensure that all spatial locations are covered by at least one block.

[0056] S5.2, the Hann window weighted fusion method is as follows.

[0057] The prediction results for each block are weighted and fused using a two-dimensional Hann window during stitching. The Hann window function w( i,j ) is defined as: , in, Indicates the ()th element within the block i,j Spatial fusion weights corresponding to ) pixel positions; , These represent the pixel's position index in the horizontal and vertical directions, respectively; Indicates the spatial dimensions of the predicted plot; Represents pi; This represents the cosine function.

[0058] The center weight of the Hann window is close to 1, while the edge weight smoothly decays to 0. The weights of adjacent overlapping blocks in the overlapping region are complementary, that is, the edge of block A is compensated by the center of block B, and vice versa.

[0059] During stitching, the weighted predictions for each spatial location are accumulated, and the weights are summed before final normalization: , Where m represents the predicted map tile number. This represents the prediction result for the m-th tile. Indicates the spatial weights of the Hann window. Indicates the spatial location of the map block.

[0060] After normalization, the image is cropped back to its original size, resulting in a seamlessly stitched large-size prediction. The smooth attenuation characteristics of the Hann window ensure the continuity of the transition between blocks and eliminate seam artifacts from direct stitching.

[0061] S5.3, inter-frame linear interpolation output.

[0062] This step is optional. If inter-frame interpolation output is specified, the following steps will be performed on the spliced ​​prediction and uncertainty respectively: 1) Output inter-frame interpolation results: retain the prediction values ​​and fill the intermediate steps with linear interpolation; 2) Insert a leading frame before the first frame: obtained by interpolation between the input end and the prediction first frame; 3) Total number of output frames: corresponding to complete and continuous day sequence prediction.

[0063] Example 2 This invention also proposes a vegetation index time series prediction system, which includes: a data preparation module and a remote sensing image inference prediction module.

[0064] The data preparation module completes the operation described in step S1 above.

[0065] In one implementation, the data preparation module performs the following operations.

[0066] Step 1: Obtain historical NDVI time series data.

[0067] Acquire NDVI time-series remote sensing images of the target area. The data comes from satellite sensors such as MODIS, Landsat, or Sentinel-2 and is stored in GeoTIFF format with dimensions of (total days, altitude, width). Each pixel value is a floating-point number in the range of [-1, 1]. Invalid observations (cloud obstruction, atmospheric interference) are marked with a specified NoData value (such as -9999.0 or -1.0) or stored directly as NaN.

[0068] Step 2: Generate training, validation, and test sets.

[0069] In one implementation, a training window is extracted from the obtained historical NDVI time series, and each window contains an input sequence and a target sequence. The input sequence is selected from the most recent (T...) in -1) × S in + 1 = (30 - 1) × 2 + 1 = 59 days in the time window, calculated by step size S in = 2 samples of T in = 30 frames of data, with input dimensions (30, H, W). The target sequence starts at an interval S after the last frame of the input sequence. out = Day 1 to Day T from day 2 pred × S out = 10 × 2 = 20 days, according to step size S out = 2 samples of T pred = 10 frames of data, with a target dimension of (10, H, W). Finally, the dataset is divided into training, validation, and test sets in a ratio of 7:1.5:1.5 and stored in different directories.

[0070] Step 3: Missing value interpolation. Perform temporal linear interpolation and spatial neighbor interpolation on the input sequence of each training sample.

[0071] Step 4: Normalization and Mask Generation. Crop the NDVI values ​​to the range of [-1, 1] to generate a valid pixel mask (valid = 1, invalid = 0), and replace NaN with -1.0 as the model input.

[0072] Step 5: Data Augmentation. Perform training sample augmentation operations randomly with a 50% probability. These operations mainly include: horizontal flipping (mirroring along the width dimension), vertical flipping (mirroring along the height dimension), and brightness adjustment (multiplying by a random factor in the range of [0.9, 1.1] and then cropping to [-1, 1]).

[0073] The remote sensing image inference and prediction module includes: encoder, decoder, jump connection module, dual-channel output head, and block weighted stitching inference module.

[0074] The encoder performs the operation described in step S2 above. Specifically, the encoder consists of three stacked ConvLSTM encoder layers, with the complete sequence of outputs from the upper layers serving as the input to the next layer. Gated convolutions are performed step-by-step, using hidden states to pass information across time steps, outputting the encoded spatiotemporal feature sequence and the final hidden states of each layer. This spatiotemporal feature sequence and the final hidden states of each layer are used by the subsequent decoder to predict future NDVI.

[0075] The three-layer ConvLSTM encoder has progressively increasing channel numbers of 64, 128, and 256. Layer 1 (64 channels) captures fine-grained local texture and edge features; Layer 2 (128 channels) models mesoscale vegetation patches and growth patterns; and Layer 3 (256 channels) extracts large-scale phenological trends and climate response features. The three ConvLSTM layers use the hidden state sequence of the lower layers as input to the upper layers, abstracting layer by layer to form a multi-scale spatiotemporal feature pyramid structure.

[0076] The encoder of this invention also includes a hidden state predictor, which consists of two 3×3 Conv layers. The input to the hidden state predictor is the complete sequence of the output of the third ConvLSTM layer, and the output is the hidden state. This hidden state is the initial state of the decoder.

[0077] The decoder also consists of three stacked ConvLSTM layers, with the number of channels decreasing sequentially to 256, 128, and 64, forming a symmetrical structure with the encoder. The decoder performs the operation described in step S3 above.

[0078] In addition, the decoder of the present invention also includes an attention module, which has channel attention and spatial attention mechanisms and is a dual-path enhancement module. The attention module performs the operation described in step S3.2 above.

[0079] Furthermore, this invention introduces a skip connection module in the feature transfer path between the encoder and decoder to fuse multi-scale spatial details and alleviate the blurring of deep feature spaces. The encoder and decoder establish skip connections using a mirror correspondence; the first, second, and third layers of the encoder are respectively connected to the third, second, and first layers of the decoder, thereby achieving the fusion of shallow spatial detail features and deep semantic features. See the steps described above for a detailed description.

[0080] The dual-channel output head performs the operation described in step S4 above. The output head performs three 3×3 convolutions on top of the last layer of features in the decoder, outputting a dual-channel prediction result: the first channel is the temporal prediction value of NDVI for each point, and the second channel is the logarithmic representation of the uncertainty for each prediction point. The mean and uncertainty of NDVI are predicted.

[0081] Block-weighted stitching inference module. Deep learning models are trained on fixed-size spatiotemporal blocks (e.g., 64×64 pixels), while practical applications require processing complete remote sensing images (typically thousands × thousands of pixels). Direct block-based inference followed by stitching leads to obvious edge effects and stitching seam problems. Therefore, this invention proposes a block-weighted stitching inference module, which implements the operation described in step S5 above.

[0082] In one implementation, the block-weighted splicing inference module performs the following operations.

[0083] Step 1: Block-wise inference.

[0084] 1) Perform block sampling on the input sequence using a sliding window strategy with a step size of 48 to obtain spatio-temporal blocks with a spatial size of 64×64. For the parts with insufficient boundary size of 64×64, the mirror padding method is adopted to extend them to the target size.

[0085] 2) For each predicted block, generate a 2D Hann window spatial weight matrix: , which is used for subsequent weighted fusion of the block prediction results.

[0086] 3) Each spatio-temporal block is inferred by the model, and a predicted mean map is output and a predicted log variance map , whose sizes are respectively , . Wherein, m represents the m th input block number, 10 represents the number of prediction time steps, and 64×64 represents the spatial size.

[0087] 4) Perform weighted fusion on the predicted mean maps of all blocks respectively and the standard deviation maps according to the 2D Hann window weights, and obtain the prediction result of the complete region through cumulative weight normalization.

[0088] 5) Crop the fused prediction result to the spatial range of the original input image (T out , H full , W full ), to obtain the final time series prediction result.

[0089] Step 2: Inter-frame interpolation output.

[0090] If the user specifies the inter-frame interpolation output mode, the following operations are performed on the spliced prediction and uncertainty respectively: 1) Output the inter-frame interpolation result: pred[t×stride] retains the predicted value, and the intermediate step s (1≤s<stride) is filled with linear interpolation; 2) Insert a leading frame before the first frame: obtained by interpolating the last input frame X last and the first prediction frame pred[0]; 3) Total number of output frames: corresponding to the complete and continuous day sequence prediction.

[0091] Step 3: NoData spatial mask post-processing.

[0092] Pixel locations where all timeframes in the input data are NaN are identified as permanent NoData regions (such as boundary-filled areas). The pixel values ​​at the corresponding locations in the prediction results are replaced with user-specified NoData flag values ​​(such as -1.0), and the corresponding locations in the uncertainty map are also marked as NoData.

[0093] Preferably, in the remote sensing image inference and prediction module, the convolutional layers are initialized using the Kaiming normal distribution, the ConvLSTM convolutional weights are initialized using orthogonal distribution, and the biases are initialized to 0. The total number of model parameters is approximately 12.2 million.

[0094] Example 3 The system of this invention is a deep learning model. Preferably, the model training hyperparameters are set as follows: Learning rate: initial learning rate is set to 0.0001, decaying according to cosine; Optimizer: Adam optimizer is used; Planned sampling decay epochs: 2 / 3 of the total training epochs; Other: batch size is set to 8, which is related to GPU memory, and the total number of training epochs is 100. The training data consists of the training set, validation set, and test set generated by the data preparation module. The forward propagation process of the module is as follows.

[0095] 1) Add channel dimensions to the input sequence (B, 30, H, W) and adjust the dimension order to (30, B, 1, H, W) to adapt to the input format of ConvLSTM.

[0096] 2) Apply a mask to the input (to filter out invalid pixels).

[0097] 3) Spatiotemporal features are extracted using a 3-layer ConvLSTM encoder to obtain the hidden state sequences of each layer.

[0098] 4) The hidden states of the encoder's last layer are mapped to the initial hidden states of the decoder via the hidden state predictor.

[0099] 5) The decoder generates predictions frame by frame (10 frames in total) through autoregression.

[0100] Preferably, a planned sampling strategy is used when training the encoder.

[0101] Specifically, to address the "exposure bias" problem caused by training-inference inconsistency, a linearly decaying planned sampling strategy is adopted. In the... In each decoding step, when the random sampling probability is τ, the actual observed value is used as the input; when the sampling probability is 1-τ, the model prediction value is used as the input. , in, Indicates the first Real-time observation of images, This indicates the model's prediction results. This indicates a channel-adjusted convolution operation. During training, the sampling probability τ is changed from the initial value τ start =1.0 is gradually reduced to τ according to a linear decay strategy. end =0.0.

[0102] , When k ≥ E decay When τ=0, the model recursively predicts using only its own predictions. k Indicates the current training round number. E decay This represents the probability decay period. This strategy enables the model to converge quickly in the early stages of training (τ is close to 1) and gradually adapt to the autoregressive conditions during inference in the later stages of training (τ is close to 0), resulting in a smooth transition and eliminating exposure bias. In the specific implementation of planned sampling, each sample within each batch is independently tossed to determine whether to use the ground truth value, and the ground truth value is only used for valid pixels (mask=1). Invalid pixels always use the model's predicted value, ensuring training consistency in missing regions.

[0103] Preferably, when training the output head, Gaussian negative log-likelihood (NLL) is used as the training target.

[0104] Specifically, assume that the prediction error of each pixel follows a mean of 0 and a variance of . If the target value follows a Gaussian distribution, then the negative log-likelihood L under this distribution is... NLL for: , in, y represents the predicted mean, and y represents the true label.

[0105] This loss function outputs a larger variance in areas with large prediction errors (such as areas with abrupt vegetation changes or cloud edges) and a smaller variance in areas with stable predictions (such as areas with uniform vegetation cover).

[0106] Figure 3 and Figure 4 Together, these findings validate the dual advantages of the proposed method: Figure 3 shows, in the time dimension, that the model-predicted sequence closely matches the actual NDVI curve, and ±1 / ±2 The uncertainty interval encompasses the true value with a reasonable coverage; Figure 4 further reveals, in the spatial dimension, the high-value region of the absolute prediction error and the standard deviation of the model output. The high-value regions exhibit significant spatial co-occurrence. These results demonstrate that the dual-channel output mechanism (mean μ and log-variance) based on Gaussian NLL end-to-end training... This endows the model with reliable cognitive self-assessment capabilities. It is not a statistical byproduct, but rather an effective proxy variable for prediction errors in both spatiotemporal dimensions. This property provides a theoretically complete basis for confidence level grading for downstream businesses: low... The region can be adopted with high confidence and directly support decision-making. The region then triggers manual verification or supplementary observations, thus systematically solving the fundamental defect of traditional deterministic prediction models that "output point estimates cannot quantify credibility".

[0107] The embodiments described above are merely preferred embodiments of the present invention. The terms "in one embodiment," "in another embodiment," "in yet another embodiment," or "in still another embodiment" used in this specification all refer to one or more of the same or different embodiments according to this disclosure. Ordinary variations and substitutions made by those skilled in the art within the scope of the present invention should be included within the protection scope of the present invention.

Claims

1. A vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding, characterized in that, include: The data preparation module acquires the historical normalized vegetation index (NDVI) time series and performs downsampling and temporal-spatial dual interpolation on the historical NDVI time series. The remote sensing image inference and prediction module includes: an encoder, a decoder, a jump connection module, and a dual-channel output head; The encoder is a three-layer stacked ConvLSTM encoder, and the decoder is a three-layer ConvLSTM autoregressive decoder structure. The encoder and decoder are connected via a skip connection module to fuse multi-scale spatial details to alleviate the blurring of deep feature spaces. Channel attention and spatial attention modules are embedded after the feature output of each decoder layer to achieve dual-path spatiotemporal feature enhancement. The dual-channel output head outputs two prediction results: the first channel is the NDVI time-series prediction value for each point, and the second channel is the logarithmic representation of the uncertainty for each prediction point.

2. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 1, characterized in that, The remote sensing image inference and prediction module also includes: The block-weighted stitching inference module achieves seamless large-area prediction based on the overlapping block-weighted stitching method using a two-dimensional Hann window. The method employs an overlapping sliding window block segmentation strategy to fill the boundary regions that are not large enough with reflections; the prediction results of each block are weighted and fused using a two-dimensional Hann window during stitching.

3. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 1, characterized in that, In the data preparation module, the temporal-spatial dual interpolation processing includes: In the first stage, linear interpolation is performed along the time axis for the temporal vector of each spatial location, and missing values ​​are filled by valid observations before and after the same location; for NaN values ​​that remain in the first stage, Euclidean distance transformation is used to find the nearest valid pixel for spatial filling.

4. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 3, characterized in that, The data preparation module also performs the following operations: Normalization and mask generation: The NDVI values ​​of the input sequence after difference are cropped to generate an effective pixel mask, and the NaN values ​​are replaced with predetermined values; Data augmentation: Randomly augment the input sequence with a 50% probability, including horizontal flipping, vertical flipping, and brightness adjustment.

5. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 1, characterized in that, The encoder performs the following operations: it performs gated convolution on the normalized vegetation index (NDVI) time series data through a three-layer ConvLSTM spatiotemporal encoder, with the hidden state transmitting information across time steps, and outputs the encoded spatiotemporal feature sequence and the final hidden state of each layer. The decoder performs the following operation: starting from the output of the autoregressive encoder at the last moment, it decodes the data step by step through a three-layer ConvLSTM spatiotemporal decoder. The skip connection module establishes a mirror correspondence between encoders. The first, second, and third layers of the spatiotemporal encoder are respectively connected to the third, second, and first layers of the autoregressive decoder to achieve the fusion of shallow spatial detail features and deep semantic features.

6. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 5, characterized in that, The attention module performs the following operations: For the high-level spatiotemporal feature sequence output by the decoder, channel attention enhancement is first used to select important features, and then spatial attention enhancement is used to select important positions, resulting in a feature map enhanced in two dimensions.

7. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 6, characterized in that, In the attention module, The feature map is globally pooled, then passed through two 1×1 convolutional layers followed by Sigmoid activation to generate channel weight vectors. Average pooling and max pooling are calculated along the channel dimension and concatenated to form a feature map for the two channels. Then, a 7×7 convolutional layer is passed followed by Sigmoid activation to generate a spatial attention map.

8. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 6, characterized in that, The decoder performs the following operations: The output features of each ConvLSTM layer in the encoder layer are aligned to the number of channels by a 1×1 convolution and then added to the decoder to form a skip connection.

9. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 2, characterized in that, The block-weighted concatenation inference module performs the following operations: 1) Adopt an overlapping sliding window block strategy: Set the block size, overlap amount and sliding step size, calculate the required number of blocks in the horizontal and vertical directions for the original image, and fill the boundary areas that are not large enough with reflection to ensure that all spatial locations are covered by at least one block. 2) Hann window weighted fusion: The prediction results of each block are weighted and fused using a two-dimensional Hann window during stitching; 3) During stitching, the weighted predictions of each spatial location are accumulated, and the weights are added together. Finally, the data is normalized and then cropped back to the original image size to obtain a large-size prediction result with seamless stitching.

10. The vegetation index time-series prediction system based on ConvLSTM spatiotemporal coding according to claim 1, characterized in that, The dual-channel output head uses Gaussian negative log-likelihood loss for probability prediction, providing a quantitative confidence estimate while outputting pixel-by-pixel predicted values.

Images