A hyperspectral image-based ground object classification model adaptive training method

Abstract

The application discloses a kind of ground feature classification model adaptive training method based on hyperspectral image, belong to the intersection technical field of remote sensing image processing and artificial intelligence.This method includes: adaptive spatial block division step, according to the training / validation / test set division of class balanced stratification with spatial non-overlapping block;Physically constrained adaptive enhancement step of unbalanced sample, impose range, smoothness and diagnostic absorption peak three layers of physical constraint to sample synthesis;Adaptive band selection step of scene perception, automatically judge scene type and adaptively determine optimal band number and combination according to information content distribution;Dataset difficulty driven model structure and hyperparameter adaptive search step, based on four-dimensional difficulty index automatically match appropriate model and hyperparameter;Optional cross-domain adaptive fine-tuning step;And dynamic early stop and failure automatic retry training scheduling execution step.The application makes overall accuracy average increase by 8.4 percentage points, minority class F1 increases by 16.4 percentage points, cross-time interval accuracy reduction amplitude reduces by 76%, GPU training time saves 40%, and can be widely applied to mineral identification, vegetation fine classification, water quality parameter inversion and other remote sensing scenes.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and artificial intelligence, specifically involving an adaptive training method for land cover classification models for hyperspectral satellite images such as Gaofen-5, environmental microsatellites, ASTER, and EO-1. It is applicable to remote sensing application scenarios such as mineral identification, fine vegetation classification, water quality parameter inversion, soil composition analysis, and fine ecological monitoring. Background Technology

[0002] Hyperspectral imagery (HSI) records the reflectance spectral characteristics of surface features through dozens to hundreds of continuously subdivided spectral bands. Compared to ordinary multispectral imagery (usually 3-10 bands), it has the advantage of "image-spectrum integration"—each pixel corresponds to a continuous spectral curve, enabling fine classification based on the diagnostic spectral absorption characteristics of ground features (such as distinguishing different crop varieties, different mineral types, and different water quality levels). Currently, mainstream commercial hyperspectral satellites include: China's Gaofen-5 (388 bands, spatial resolution 30m, full VNIR / SWIR spectrum), the HJ-1A HSI environmental microsatellite (115 bands, 100m), the US ASTER (14 bands, 15-90m), and the US EO-1 Hyperion (242 bands, 30m), etc.

[0003] Land cover classification is a core technical aspect of hyperspectral image applications. Traditional methods include: ① spectral matching-based methods (spectral angle matching (SAM), spectral information divergence (SID), matched filtering (MF); ② machine learning-based methods (Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN); ③ deep learning-based methods (one-dimensional CNN, two-dimensional CNN, three-dimensional CNN, Transformer, Generative Adversarial Network (GAN) semi-supervised learning).

[0004] Existing technical solutions mainly include the following three categories: The first category is supervised training methods based on fixed labeled samples. The standard process of mainstream commercial software (ENVI, ERDASIMAGINE) and open source tools is as follows: after the user manually labels a small number of samples on the hyperspectral image, they are randomly divided into training set and validation set in a 7:3 ratio. The model structure and hyperparameters are fixed. The model is trained on the training set and monitored for early stopping by the validation set. Finally, the classification results are output on the test set.

[0005] The second category is preprocessing and training methods based on fixed band selection. Before training, band selection is performed to remove redundant bands, and then the classification model is trained. Representative methods include band selection based on mutual information (MIFS), band selection based on discrete particle swarm optimization (DPSO-BS), adaptive band selection based on attention mechanism (ABSA), or fixed selection of the first N bands with the largest information content, and subsequent data are trained and classified using the same subset of bands.

[0006] The third category is semi-supervised learning plus data augmentation methods, which generate pseudo-labeled samples through generative adversarial networks (GANs) or use contrastive learning methods such as FixMatch / SimCLR to train with a large number of unlabeled samples to reduce the reliance on manual annotation.

[0007] The aforementioned existing technical solutions have the following six major drawbacks in actual commercial projects: Defect 1: Training set distribution shift caused by fixed sample partitioning strategy. The existing solution, "random 7:3 partitioning by pixels," ignores the spatial continuity of hyperspectral imagery. Adjacent pixels are highly correlated spectrally, and random partitioning leads to spectral leakage in the training and validation / test sets. The accuracy of the validation set is artificially inflated by 5-8 percentage points, but when actually applied to the next image from the same satellite, the accuracy drops sharply by 5-12 percentage points, resulting in customer rejection.

[0008] Defect 2: Failure to handle sample class imbalance. Hyperspectral imagery is naturally imbalanced in terms of ground features. Existing methods such as weighted random oversampling (e.g., SMOTE) generate invalid synthetic samples in the 100+ dimensional hyperspectral space. If the synthetic spectral curves violate physical laws (reflectance greater than 1 or less than 0), it will actually reduce the classification accuracy of rare ground feature categories.

[0009] Defect 3: Band selection relies on manual experience or fixed thresholds, making it unable to adapt to different scenarios. Existing methods fix the number of bands (e.g., selecting the first 50 bands), but the information distribution varies greatly across different application scenarios. Vegetation classification has high information density in the visible and near-infrared bands, while mineral identification has strong diagnostic features in the short-wave infrared band. Fixed thresholds lead to high redundancy or omission of key diagnostic bands, making it impossible to reuse across scenarios. Each new project requires algorithm engineers to manually adjust the band selection parameters.

[0010] Defect 4: The model structure and hyperparameters are hard-coded and cannot be adaptively adjusted according to the difficulty of the data. The existing code hardcodes the model structure and training hyperparameters into the configuration file. In simple scenarios, the model is too large and overfits, while in complex scenarios, the model is too small and underfits. Manually tuning the hyperparameters takes an average of 2-3 days to iterate through multiple parameter combinations for each dataset.

[0011] Defect 5: Lack of cross-image domain adaptability, with accuracy plummeting across different time phases or different sensors. Existing solutions, after training on a single image, are directly used on images of the same region from the same satellite at different time phases. Due to changes in vegetation phenology and atmospheric conditions, the overall accuracy drops by more than 18%; across sensors, the accuracy drops even further to below 65%. Images of the same region at different time phases or from different satellites must be re-annotated and trained, doubling the procurement cost.

[0012] Defect 6: Lack of adaptive early stopping and retry mechanisms during training. Existing solutions fix the number of training epochs or the early stopping tolerance value on the validation set, resulting in more than 5 times the GPU time being wasted on simple datasets, oscillating and stopping early on difficult datasets to reach local optima, and training diverging when the random seed is poor but still completing all epochs. In the hyperspectral training process, 32% of the GPU time is wasted on ineffective training.

[0013] Therefore, in order to address the above six major shortcomings, there is an urgent need for a hyperspectral image land cover classification model training method that can adaptively solve the problems of the entire process of sample partitioning, sample enhancement, band selection, model matching, cross-domain transfer and training scheduling. Summary of the Invention

[0014] (a) Technical problems to be solved The purpose of this invention is to address the six major shortcomings of existing technologies by providing an adaptive training method and system for land cover classification models based on hyperspectral imagery, thereby solving the following technical problems: 1. Resolve the issues of spectral leakage in the training / test sets and artificially high validation accuracy caused by "random pixel 7:3 partitioning"; 2. Solve the problem of physically invalid spectral curves generated by traditional oversampling when there is class imbalance in high-dimensional space; 3. The band selection method adaptively selects the number and combination of bands based on the differences in spectral information distribution in the application scenario, rather than using a fixed threshold; 4. The model structure complexity and training hyperparameters are automatically adjusted according to the difficulty of the dataset, eliminating the need for manual iteration of parameter combinations; 5. After training with a small number of labeled source images, it can adaptively transfer to target images of different regions at the same time, different time phases in the same region, or even different sensors, with the accuracy loss within an acceptable range. 6. Implement a dynamic early stopping and automatic retry mechanism during the training process to save GPU computing time to the maximum extent while ensuring accuracy.

[0015] (II) Technical Solution To address the aforementioned technical problems, this invention proposes an end-to-end adaptive training framework consisting of "6 major adaptive modules plus 1 unified scheduler," and the overall technical solution is as follows: An adaptive training method for a land cover classification model based on hyperspectral imagery includes the following steps: S1. Adaptive spatial block partitioning steps: Read the hyperspectral image and user-annotated sample set, adaptively calculate the block side length according to the image size and sample number, divide the image into non-overlapping spatial blocks, and use the class-balanced hierarchical allocation algorithm to allocate the spatial blocks into training set, validation set and test set, ensuring that the training block and the validation / test block are completely non-adjacent in space. S2, Physically Constrained Imbalanced Sample Adaptive Augmentation Steps: Count the number of samples of each class in the training set, identify the minority class set, perform synthetic augmentation on the minority class samples, and apply three layers of physical constraint tests: value range constraint, smoothness constraint, and diagnostic absorption peak constraint. Only retain the synthetic samples that pass all tests, and output the class-balanced augmented training set. S3. Scene-aware adaptive band selection steps: Calculate the comprehensive information content score of each band, automatically determine the application scene type and adjust the weight of the scene-specific spectral index accordingly, and adaptively determine the optimal number of bands and band index set based on the information content accumulation ratio and sample size. S4. Model structure and hyperparameter adaptive search steps driven by dataset difficulty: Calculate the four-dimensional comprehensive difficulty score of the augmented training set, including the difficulty of the number of classes, imbalance difficulty, spectral noise difficulty, and sample size difficulty. Match candidate model structures in the predefined model pool according to the difficulty level, and automatically search for the optimal training hyperparameters under the Bayesian optimization framework. S5. Training scheduling execution steps for dynamic early stopping and automatic retry of failure: Adaptively set the maximum number of training rounds, early stopping threshold, oscillation detection threshold and divergence detection threshold according to the comprehensive difficulty score. After each round of training, three types of detection are executed in parallel: normal convergence early stopping detection, training oscillation learning rate reduction recovery detection and training divergence automatic retry detection. The final trained ground cover classification model is output.

[0016] To elaborate further, the above method may also include an optional cross-domain adaptive fine-tuning step: after training a pre-trained model with labeled images in the source domain, statistical alignment based on spectral physical consistency is first performed on unlabeled images in the target domain, and then semi-supervised fine-tuning based on consistency regularization is performed to achieve cross-temporal or cross-sensor domain adaptation.

[0017] Correspondingly, the present invention also provides an adaptive training system for a land cover classification model based on hyperspectral imagery, including: an adaptive spatial block partitioning module, a physically constrained imbalanced sample adaptive enhancement module, a scene-aware adaptive band selection module, a dataset difficulty-driven adaptive search module for model structure and hyperparameters, a training scheduling module for dynamic early stopping and automatic retrying of failures, and a unified scheduling interface module for coordinating the execution order and data transmission of each module; it may also optionally include a cross-domain adaptive fine-tuning module.

[0018] The specific technical solutions for the six modules are explained in detail below:

[0019] This module improves the traditional "random partitioning by pixel" to "hierarchical partitioning by spatial block level plus class balance", which ensures that the training blocks and validation blocks do not overlap in space, and the validation accuracy is truly reliable.

[0020] The specific steps are as follows: Step S1-1: Read the user-annotated sample set and record the spatial coordinates (x, y) and category c for each sample pixel.

[0021] Step S1-2: Calculate the spatial distribution density of the labeled samples using an adaptive grid side length s: s = max(32, floor(sqrt(W×H / K_block))) Where K_block is the number of target blocks, adaptively set to max(16, ceil(N_sample / 50)), N_sample is the total number of sample pixels, and W and H are the image width and height, respectively.

[0022] Step S1-3: Divide the entire image into a set of non-overlapping blocks according to an s×s grid.

[0023] Steps S1-4: Count the number of labeled samples and the number of categories covered in each block, and sort all blocks from largest to smallest by the product of the number of samples and the number of categories.

[0024] Step S1-5: Use the class-balanced hierarchical allocation algorithm to allocate blocks to the training set (60% blocks), validation set (20% blocks), and test set (20% blocks), ensuring that: the proportion of each class in the training / validation / test set deviates from the global class proportion by less than 5%; and there is at least one block buffer interval between the training block and the validation / test block (they are not spatially adjacent).

[0025] Output the final partitioning result, and record the partition block number and coordinate range.

[0026] 2. Physically Constrained Imbalanced Sample Adaptive Augmentation Module (PCAA) This module adds three layers of physical constraints to the traditional oversampling to ensure that the synthesized enhanced samples are physically legitimate spectral curves of ground features.

[0027] The specific steps are as follows: Step S2-1: Count the number of samples n_c (c=1..N) of each class in the training set, and find the minority class set with the fewest samples C_minor = {c : n_c < α·max(n_c)}, where α is the imbalance ratio threshold, which is adaptively set to 0.3.

[0028] Step S2-2: For each minority class c, the number of synthesized sample augmentations aug_c = β·(max(n_c) -n_c), where β is the adaptive augmentation ratio, which is set to 0.8 to avoid overfitting caused by completely flattening to the maximum number of classes.

[0029] Step S2-3: For each minority class sample pair (x_i, x_j) that needs to be synthesized, perform weighted linear interpolation in the spectral dimension to obtain the initial synthesized sample x' = x_i + λ·(x_j - x_i), where λ follows a uniform distribution U(0,1).

[0030] Step S2-4: Apply three layers of physical constraint checks; only if all three pass are the synthetic sample retained. (1) Value range constraint: The reflectance values ​​of all bands must be within the dynamic range of the sensor (e.g., [0,1] or conform to the DN value range of the satellite), otherwise they will be discarded.

[0031] (2) Smoothness constraint: The sum of the absolute values ​​of the first-order differences of the synthetic sample spectral curves is less than γ times the average first-order difference of the original samples of the same type, where γ is 1.5; if the violation is violated, the Savitzky-Golay smoothing kernel (window size 5, polynomial order 2) is used to perform a spectral smoothing test on x', and if the violation is still violated, it is discarded.

[0032] (3) Diagnostic absorption peak constraints: According to the prior spectral knowledge base of land cover category c, x' must have a prior absorption characteristic (local minimum) in the corresponding band interval. For example, vegetation must have an absorption valley at the 680nm red edge position, and kaolinite minerals must have an Al-OH absorption peak near 2200nm; if it is violated, it will be discarded.

[0033] Step S2-5: If the effective synthesis rate of a certain type of sample is less than 30% after 3 layers of constraints, then Gaussian perturbation enhancement of the spectral feature space is used instead of linear interpolation: x' = x_i + N(0, σ_c^2·I), where σ_c is 0.05 times the standard deviation of each band of the c-th type of sample; after Gaussian perturbation, the above 3 layers of physical constraints are also checked. Output the physically valid class-balanced training set.

[0034] 3. Scene-Aware Adaptive Band Selection Module (SA-ABS) This module first automatically determines the application scenario type, and then adaptively determines the number of bands K and the specific band index based on the distribution of band information.

[0035] The specific steps are as follows: Step S3-1: Calculate the comprehensive information content score for all B bands, integrating three evaluation indicators (with adaptive weights): Score(b) = w1·I(b) + w2·E(b) + w3·RSI(b) Where I(b) is the information entropy of band b (the richness of pixel value distribution, a general indicator with high basic weight); E(b) is the reciprocal of the mean mutual information between band b and adjacent bands b-1 and b+1 (a redundancy indicator, the higher the mutual information, the more redundant); RSI(b) is the relative spectral index contribution of band b, which is adaptive according to the scene type.

[0036] Step S3-2: Automatic scene determination (no need for user to manually input scene type): (1) Calculate the average first derivative value D_red of the average spectral curve of each class in the training sample at the red edge position of 680nm. If |D_red| is greater than the first threshold θ1, it is judged as a vegetation-dominated scene. RSI(b) is the sensitivity of the vegetation spectral index (NDVI / EVI / MTCI, etc.) to band b, and w3 is 0.4.

[0037] (2) Calculate the absorption depth A_swir of the average spectrum of each class in the training sample in the two short-wave infrared absorption feature regions of 2200nm / 2350nm. If A_swir is greater than the second threshold θ2, it is judged to be a mineral-dominated scene. RSI(b) is taken as the absorption feature of mineral diagnosis (such as the absorption intensity of kaolinite / alumite / montmorillonite in the corresponding band in the USGS spectral library), and w3 is taken as 0.45.

[0038] (3) Statistically calculate the reflectance ratio of various types in the 550-700nm visible light region and the 750-900nm near-infrared region in the training samples. Clear water bodies show a sharp drop in near-infrared reflectance. If this characteristic is met, it is judged to be a water quality-dominated scenario. RSI(b) takes the sensitive band weight of water quality parameters (chlorophyll a / total suspended matter / turbidity), and w3 is 0.35.

[0039] (4) If none of the above conditions are met, it is judged as a general scenario, and all RSI(b) are set to 1 (equal weight), and w3 is set to 0.2.

[0040] The sum of the weights w1, w2, and w3 is normalized to 1.

[0041] Step S3-3: Sort all bands in descending order of the composite score(b).

[0042] Step S3-4: Adaptively determine the optimal number of bands K: (1) The information content scores after step-by-step accumulation and sorting are calculated as the proportion of the accumulated score to the total score: Cum(K) = Σ(i=1 to K) Score(i) / Σ(i=1 to B) Score(i).

[0043] (2) Find the smallest K that satisfies Cum(K) ≥ T, and the threshold T is adaptively set as follows: for small sample scenarios (<30 samples per class), T=0.90, and for large sample scenarios, T=0.95.

[0044] (3) Simultaneously apply upper and lower bound constraints to K: K ∈ [max(10, floor(B / 10)), min(80, floor(B / 2))] to avoid selecting too few or too many bands.

[0045] Step S3-5: Output the final set of K selected band indices, and also output the automatically determined scene type labels and score curves for each band.

[0046] 4. Dataset Difficulty-Driven Model Structure and Hyperparameter Adaptive Search Module (DDANAS) This module first calculates the 4-dimensional difficulty index of the dataset, and then automatically selects the optimal model structure and the optimal hyperparameter combination from the predefined candidate model pool based on this index.

[0047] The specific steps are as follows: Step S4-1: Calculate the 4-dimensional difficulty index of the augmented training set: (1) D1 (Number of categories difficulty): The number of categories N is normalized to [0,1], and 1 is taken for N≥12.

[0048] (2) D2 (unbalanced difficulty): Gini coefficient of number of categories G = 1 - Σ(p_c^2), where p_c is the proportion of each category. The more unbalanced the category, the smaller G is. The difficulty is normalized after taking 1-G.

[0049] (3) D3 (Spectral noise difficulty): Randomly select 1000 pixels to perform Savitzky-Golay spectral smoothing, calculate the mean square error (MSE) before and after smoothing. The larger the MSE, the higher the noise. Normalize.

[0050] (4) D4 (sample size difficulty): the average number of samples per class avg(N_c). The smaller the sample size, the higher the difficulty. Normalize by 1 - min(1,avg(N_c) / 300).

[0051] Overall difficulty score: D = 0.3·D1 + 0.3·D2 + 0.2·D3 + 0.2·D4, D ∈ [0,1].

[0052] Step S4-2: Automatically select 1-3 candidate model structures from the predefined model pool based on the difficulty D: (1) 0 ≤ D < 0.3 (simple scenario): three-layer one-dimensional convolutional neural network (64 channels, about 50,000 parameters) and radial basis kernel support vector machine (C=1, γ=auto). (2) 0.3 ≤ D < 0.6 (Medium difficulty): Four-layer three-dimensional convolutional neural network (3×3×3 convolutional kernel, about 500,000 parameters) and HybridSN hybrid network (two-dimensional CNN plus three-dimensional CNN hybrid, about 800,000 parameters). (3) 0.6 ≤ D ≤ 1.0 (difficult scenario): SpectralFormer lightweight Transformer (2 layers, 4 heads, about 3 million parameters) and SSFTT spatial spectral fusion Transformer (about 5 million parameters).

[0053] Step S4-3: For each candidate model, automatically search for the optimal training hyperparameters within the Bayesian optimization framework. The search space includes: learning rate (log-uniformly sampled at [1e-5, 5e-3]), weight decay coefficient (log-uniformly sampled at [1e-6, 1e-3]), batch size (sampled from {16, 32, 64, 128, 256} with upper and lower bounds determined by sample size), optimizer type (sampled from {AdamW, SGD_momentum0.9}), and data augmentation probability (uniformly sampled at [0, 0.6], increased for difficult scenarios). The number of Bayesian optimization trials is adaptive: 40 trials for high difficulty D and 12 trials for low difficulty D, with each trial running for only 15 rounds to quickly verify the approximate effect.

[0054] Step S4-4: After Bayesian optimization, select the model structure with the highest Kappa coefficient on the validation set and add hyperparameters as the final optimal solution.

[0055] 5. Cross-domain adaptive fine-tuning module (CDAFT, optional) This module performs two-stage domain adaptation fine-tuning for the target domain (unlabeled) based on training in the source domain (labeled), using unlabeled data from the target domain to reduce domain differences.

[0056] The specific steps are as follows: Step S5-1: Perform normal pre-training on the source domain data using the process of steps S1-S4 to obtain the pre-trained model.

[0057] Step S5-2: The first stage is statistical alignment based on spectral physical consistency, which does not require training and only involves linear transformation of the feature space. (1) Calculate the mean μ_{s,b} and standard deviation σ_{s,b} of each band b in the source domain, and the mean μ_{t,b} and standard deviation σ_{t,b} of the corresponding band b in the unlabeled target domain; (2) Perform linear normalization alignment on band b of each pixel in the target domain: H'_{t,b} = ((H_{t,b} - μ_{t,b}) / σ_{t,b})·σ_{s,b} + μ_{s,b}; (3) At the same time, the spectral curve of the target domain after alignment still satisfies the physical constraints (range plus smoothness, constraints 1 and 2 of the same module 2), to avoid illegal spectra after alignment.

[0058] Step S5-3: Second stage, semi-supervised fine-tuning based on consistency regularization: (1) Freeze the first 70% of the bottom-level feature extractors of the pre-trained model and only fine-tune the last 30% of the top-level classifiers to avoid catastrophic forgetting; (2) Perform two different data augmentations on each unlabeled pixel in the target domain (randomly add Gaussian noise σ=0.01 and random band dropout ratio 0.05) to obtain two predictions p1 and p2; (3) Training loss = cross-entropy loss of labeled samples in the source domain L_CE + λ·consistency loss of unlabeled samples in the target domain L_con = MSE(p1, p2), where λ is adaptively selected: λ=0.3 when the same satellite is used across time, and λ=0.8 when the sensor is used across time. (4) The number of fine-tuning rounds is fixed at 20 rounds to avoid overfitting to the unlabeled pseudo-label noise of the target domain.

[0059] Step S5-4: Output the adapted model and directly classify the target domain image.

[0060] 6. Dynamic Early Stopping and Automatic Retry Training Scheduling Module (DESAR) This module dynamically monitors three types of signals during the training process and automatically decides whether to continue training, stop the convergence early, or retry immediately.

[0061] The specific steps are as follows: Step S6-1: Before each training session, set the training budget parameters according to the difficulty level D: Maximum number of rounds 40 80 150 Early cessation tolerance value p 5 10 15 Shock detection tolerance value s 3 5 8 Divergence detection threshold The validation set loss is greater than 2 × the initial loss for three consecutive rounds. 5 consecutive rounds 8 consecutive rounds Maximum number of automatic retries 1 time 2 times 3 times Step S6-2: After each round of training, perform three types of detection in parallel: (1) Early stopping detection of normal convergence: If the Kappa coefficient of the validation set does not increase for p consecutive rounds (increase threshold < 0.002), early stopping is triggered, and the model weight of the optimal round is saved.

[0062] (2) Training Oscillation / Local Optimality Detection: If the Kappa coefficient of the validation set fluctuates within ±0.01 for s consecutive rounds, and the current Kappa is less than 0.9 times the baseline Kappa of the same model on the same difficulty dataset in the past, the learning rate will be automatically reduced to 1 / 3 of the original, and training will be restarted from the current optimal weight. The learning rate will be reduced a maximum of 2 times.

[0063] (3) Training divergence / poor random seed detection: If the training loss is greater than the divergence detection threshold, or the validation set Kappa < 0.2 and there is no upward trend in the first 5 rounds, the training is terminated immediately, the random seed is automatically replaced and the next suboptimal combination of hyperparameters in module 4 is restarted, and the training is retried.

[0064] Step S6-3: If the minimum acceptable Kappa threshold (adaptive threshold = Kappa of the SVM method on the same dataset + 0.03) is not reached after all retries have been exhausted, an alarm will be triggered and a diagnostic log will be output, and manual intervention will be initiated for further investigation.

[0065] Step S6-4: Output the optimal model weights plus the complete training log.

[0066] (III) Beneficial Effects Compared with existing solutions (ENVI default SVM plus 7:3 random division plus fixed 50 band selection) in actual projects, the beneficial effects of this invention are as follows: 1. The overall accuracy (OA) of the final test set improved from 85.3% to 93.7%, an increase of 8.4 percentage points; 2. The F1 value for minority categories (rare minerals / rare crops) increased from 64.8% to 81.2%, an increase of 16.4 percentage points; 3. The decrease in OA (Out-of-Home) for cross-temporal imagery decreased from 18.2% to 4.3%, a reduction of 76%; 4. Algorithm engineers' parameter tuning time has been reduced from 2.6 days per project to 0 days (fully automated), saving 100% of manpower and time; 5. GPU training time (per A10 single card per project) reduced from 4.2 hours to 2.5 hours, a 40% reduction in GPU time; 6. The project delivery cycle was shortened from 10.3 days to 5.1 days, a 50% reduction in delivery time; 7. The rework rate due to accuracy issues for customers dropped from 27% to 0%, and rework was completely eliminated.

[0067] The aforementioned beneficial effects are all derived from real comparative data from actual projects and can be further demonstrated through specific embodiments and ablation experiments in this specification. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the overall framework of the present invention, showing the end-to-end process of six adaptive modules plus a unified scheduler; Figure 2 This is a comparative diagram of the spatial block division of Module 1 and the existing random pixel division. The left diagram shows the case where random 7:3 red and blue pixels are mixed together, and the right diagram shows the case where the spatial block division is performed and the red and blue blocks are not adjacent. Figure 3The flowchart for the three-layer physical constraint verification of module 2 shows the branch judgment logic of discarding, retaining, and smoothing. Figure 4 The flowchart for the automatic scene judgment in Module 3 shows four branches: red edge judgment, shortwave infrared judgment, water ratio judgment, and general judgment. Figure 5 A bar chart comparing the distribution of selected bands for three typical scenarios (vegetation, minerals, and water quality) in Module 3. The horizontal axis represents the spectral wavelength, and the vertical axis represents the distribution frequency of the selected bands. Figure 6 The 4D difficulty radar chart for Module 4 shows the difficulty comparison of the datasets of three specific embodiments on the four axes D1-D4; Figure 7 The diagram shows the mapping relationship between Module 4 difficulty D and the corresponding recommended model pool, illustrating the three difficulty ranges and their corresponding model cards. Figure 8 The flowchart for the two-stage cross-domain fine-tuning of Module 5 shows the process of statistical alignment, consistency fine-tuning, output, and the ratio of frozen / trained layers. Figure 9 The state machine diagram for the three types of detection mechanisms in Module 6 shows the state transition relationships for normal training, early stopping, oscillating learning rate reduction, and divergent retry. Figure 10 This is a comparison chart of the classification results of the Tangshan, Hebei project in Example 1. The left image is a true-color image, the middle image is the classification result of the existing SVM method, and the right image is the classification result of this method. Figure 11 The image shows a magnified comparison of the classification results of the rare mineral pyrophyllite in Example 1. The left image shows the case where SVM missed detection, and the right image shows the case where the corresponding patch was completely extracted by this method. Figure 12 The ablation experiment results are grouped into bar charts, showing three sets of comparative data: baseline method, adding each module one by one, up to the full module, OA, minority class F1, and GPU time. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0070] Example 1: Mineral Identification Project Based on Gaofen-5 Sediment in a Mining Area This implementation scenario involves a bauxite mining area, requiring the identification of seven types of features: background bare land, vegetation, water bodies, alunite, kaolinite, pyrophyllite, and montmorillonite. The client requires an F1 score of ≥80% for the latter four rare mineral types.

[0071] The hyperspectral data used in this embodiment is imagery from the Gaofen-5 satellite, acquired on April 12, 2025, covering an area of ​​approximately 240 km², with a band count of B=296 (VNIR 150 band plus SWIR 146 band; after removing bad bands, 274 bands are actually used), and a spatial resolution of 30 m. The labeled sample contains 4382 pixels, distributed across seven categories: bare background 1620 (37%), vegetation 1158 (26%), water 89 (2%), alunite 487 (11%), kaolinite 402 (9%), pyrophyllite 362 (8%), and montmorillonite 264 (6%), classifying it as a highly imbalanced dataset.

[0072] Step S1 (Adaptive Spatial Block Partitioning) Execution Results: Image size W=900, H=889, number of samples 4382, adaptive block side length s=64 according to the formula, number of blocks K_block=196; after class-balanced stratification 6:2:2 allocation, the number of training / validation / test samples are 2675 / 859 / 848 respectively, the deviation of each category ratio is less than 3.2%, which meets the ≤5% requirement, and the training block and the validation / test block are at least 1 block apart.

[0073] Step S2 (physical constraint sample enhancement) execution results: Imbalance threshold α=0.3, minority classes identified as water (2%), pyrophyllite (8%), and montmorillonite (6%); enhancement ratio β=0.8, a total of 1210 enhanced samples need to be synthesized; the pass rate of the 3-layer physical constraint method is: 48% pass by linear interpolation method (52% were discarded due to violation of smoothness or absorption peaks), and 76% pass by Gaussian perturbation alternative scheme; the final enhanced class ratio deviation is ≤11% (original maximum deviation is 35%), and the spectral curves of all synthesized samples are in accordance with physical laws after manual inspection by experts.

[0074] Step S3 (Adaptive Band Selection) Execution Results: During automatic scene judgment, the absorption depth A_swir=0.041 of shortwave infrared 2200 / 2350nm is greater than the second threshold θ2=0.02, and it is judged as a mineral-dominated scene with weights w1=0.3, w2=0.25, and w3=0.45. The sample size is moderate, the information accumulation threshold T=0.93, and the optimal number of bands K=42 is adaptively determined (far lower than the commonly used fixed 80 bands, saving 47.5% of the computation). Among the 42 selected bands, 31 are located in the SWIR band range (2000-2500nm), which is consistent with the physical prior of the mineral diagnostic characteristic distribution.

[0075] Step S4 (Difficulty-Driven Model Search) execution results: The 4-dimensional difficulty indexes are D1 (7 classes) = 0.58, D2 (imbalance) = 0.76, D3 (noise) = 0.35, and D4 (average samples 626) = 0.0, with an overall difficulty of D = 0.51 (slightly above average). Two candidate models, HybridSN and 3D-CNN-4 layers, were matched. Bayesian optimization was performed 25 times (15 rounds of fast validation each time), and the optimal result was HybridSN with a learning rate of 3.7e-4, weight decay of 2.1e-5, batch size of 64, AdamW optimizer, and boost probability of 0.32.

[0076] Step S5 (Dynamic Training Scheduling) execution result: D=0.51 corresponds to medium difficulty parameters: maximum number of rounds 80, early stop tolerance value 10, oscillation detection 5, divergence detection 5 rounds, 2 retries; in actual training, the validation set Kappa was triggered to stop early after 10 consecutive periods without improvement in the 43rd round, and the optimal round was the 33rd round; the actual GPU time was 48 minutes, compared with 1 hour and 52 minutes for 100 rounds with fixed parameters manually, saving 57% of the time.

[0077] Optional cross-domain step execution results: The client additionally requires another image from the same satellite in September 2025 (different time phase) of the same mining area; after statistical alignment and semi-supervised fine-tuning, the cross-domain classification OA improved from 74.6% using the source model directly to 89.8%, compared to the source model's own image test OA of 94.2%, a decrease of only 4.5%, meeting the contract requirement of ≤10%.

[0078] The final accuracy comparison of this embodiment is as follows: Overall OA accuracy 83.1% 94.2% +11.1% AA average accuracy 71.8% 87.5% +15.7% Kappa 0.783 0.927 +0.144 Alunite F1 72.3% 86.8% +14.5% Kaolinite F1 68.1% 83.4% +15.3% Pyrophyllite F1 62.4% 82.1% +19.7% (Contracts ≥80% compliant) Montmorillonite F1 59.3% 80.5% +21.2% (Contract compliance rate ≥ 80%) The existing ENVI solution fails to meet the F1 standard for all four rare minerals and requires rework; the method of this invention achieves all the standards on the first attempt.

[0079] Example 2: Agricultural Fine Classification Project of Gaofen-5 Satellite in a Certain Region This implementation scenario involves a major grain-producing county, requiring the identification of 10 types of crops: winter wheat, corn, soybeans, cotton, peanuts, orchards, vegetable greenhouses, water bodies, buildings, and bare land. The client requires an F1 score of ≥85% for winter wheat / corn / soybeans, used for planting area statistics and subsidy distribution.

[0080] The data is from Gaofen-5 imagery taken on June 18, 2025, with 274 effective bands, covering an area of ​​approximately 3200 km², and 8920 labeled pixels, distributed relatively evenly across 10 categories.

[0081] Key execution results: Module 1 adaptive block side length s=128, K_block=256; after 6:2:2 partitioning, the verification accuracy showed no artificially high value (difference from the test set <1.2%).

[0082] Module 3 automatically identifies the scene as vegetation-dominated (red edge feature d_red=0.0048 is greater than the first threshold θ1=0.003), and adaptively selects K=37 bands, of which VNIR vegetation-sensitive bands account for 78%.

[0083] Module 4 has an overall difficulty of D=0.39 (10 classes but relatively balanced, low noise). The HybridSN model was selected, and Bayesian optimization was performed 18 times for rapid validation.

[0084] Module 6 actually stopped early 37 times, with the GPU taking 1 hour and 22 minutes, compared to 2 hours and 58 minutes for the existing fixed solution, saving 54%.

[0085] Final accuracy comparison: OA improved from 86.9% to 94.8% (+7.9%), Kappa improved from 0.847 to 0.941 (+0.094), winter wheat F1 improved from 82.1% to 91.2% (+9.1%), corn F1 improved from 79.4% to 89.3% (+9.9%), and soybean F1 improved from 74.8% to 86.7% (+11.9%, contract ≥85% met).

[0086] Example 3: ASTER Ecological Assessment Project in a Nature Reserve (Cross-Domain Validation) This implementation scenario is a national nature reserve. The ecological assessment requires classification of 9 land cover categories (evergreen broad-leaved forest, coniferous forest, bamboo forest, grassland, shrubland, water body, bare rock, road, and man-made structures). The source domain is ASTER 2023 imagery, manually labeled at 5600 pixels; the target domain is Gaofen-5 2025 imagery of the same reserve, unlabeled, requiring an OA ≥ 85% after cross-sensor and cross-temporal domain shift.

[0087] The key role of cross-domain module 5: Existing solutions directly apply the model trained by ASTER to Gaofen-5 imagery with an OA of only 62.3%; when only alignment is statistically analyzed, the OA is 78.1%; when only consistency fine-tuning is performed, the OA is 83.6%; the two-stage joint approach of this invention (first statistical alignment λ, then statistical matching alignment, followed by consistency fine-tuning with λ=0.8) ultimately achieves a cross-domain OA of 87.9%, meeting the customer's requirement of ≥85%, without requiring any new annotations to the target image, saving the customer approximately RMB 28,000 in third-party annotation costs.

[0088] To further verify the independent contribution of each module in this invention, ablation experiments were conducted. Using a baseline of 3D-CNN-4 layers with 7:3 random partitioning, no enhancement, fixed 50 bands, manually fixed superpartitions (no cross-domain involvement), and fixed 100 rounds, each module was added one by one, and the results are as follows: Baseline (existing scheme) 85.3% 64.8% 1.00× 18.2% Add Module 1 space block partitioning 87.2%（+1.9%） 67.5% 1.00× 15.4% Add Module 2 Physical Constraint Enhancement 89.6%（+2.4%） 75.1% (+7.6% maximum) 1.05× 14.1% Add Module 3 Adaptive Band Selection 90.8%（+1.2%） 76.8% 0.78 × (22% faster) 13.2% Add Module 4 Difficulty Driver for NAS 91.5%（+0.7%） 78.3% 0.70× 12.5% Add Module 5 for cross-domain fine-tuning 92.3%（+0.8%） 79.4% 0.73× 4.3% (Maximum Improvement) Add Module 6 for dynamic early stop scheduling 92.4% (lossless accuracy) 79.6% 0.59 × (Total 41%) 4.3% Full module (complete method of this invention) 93.7% 81.2% 0.59× 4.3% Ablation experiments show that each of the six modules independently contributes quantifiable performance improvements or time savings, and each module is an indispensable technical feature.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. For example, the scene judgment of module 3 can be expanded to include urban scene judgment, the model pool of module 4 can be expanded to include SpectralGPT large models, module 5 can be expanded to include cross-resolution migration, and module 6 can be expanded to include automatic model degradation after multiple failures, etc. These variations are all within the scope of protection of the claims of the present invention.

Claims

1. An adaptive training method for a land cover classification model based on hyperspectral imagery, comprising the following steps: S1. Adaptive spatial block partitioning steps: Read the hyperspectral image and user-annotated sample set, adaptively calculate the block side length according to the image size and sample number, divide the image into non-overlapping spatial blocks, and use the class-balanced hierarchical allocation algorithm to allocate the spatial blocks to the training set, validation set and test set in a 6:2:2 ratio to ensure that the training block and the validation / test block are completely non-adjacent in space. S2, Physically Constrained Imbalanced Sample Adaptive Augmentation Steps: Count the number of samples of each class in the training set, identify the minority class set, perform synthetic augmentation on the minority class samples, and apply three layers of physical constraint tests: value range constraint, smoothness constraint, and diagnostic absorption peak constraint. Only synthetic samples that pass all tests are retained. When the effective synthesis rate is less than 30%, the Gaussian perturbation augmentation scheme is automatically switched to, and a class-balanced augmented training set is output. S3. Scene-aware adaptive band selection steps: Calculate the comprehensive information content score of each band, automatically determine the application scene type and adjust the weight of the scene-specific spectral index accordingly, and adaptively determine the optimal number of bands and band index set based on the information content accumulation ratio and sample size. S4. Model structure and hyperparameter adaptive search steps driven by dataset difficulty: Calculate the four-dimensional comprehensive difficulty score of the augmented training set, including the difficulty of the number of classes, imbalance difficulty, spectral noise difficulty, and sample size difficulty. Match candidate model structures in the predefined model pool according to the difficulty level, and automatically search for the optimal training hyperparameters under the Bayesian optimization framework. S5. Training scheduling execution steps for dynamic early stopping and automatic retry of failure: Adaptively set the maximum number of training rounds, early stopping threshold, oscillation detection threshold and divergence detection threshold according to the comprehensive difficulty score. After each round of training, three types of detection are executed in parallel: normal convergence early stopping detection, training oscillation learning rate reduction recovery detection and training divergence automatic retry detection. The final trained ground cover classification model is output.

2. The method according to claim 1, characterized in that, The formula for adaptively calculating the block side length in step S1 is: s = max(32, floor(sqrt(W×H / K_block))) Where K_block is the number of target blocks, adaptively set to max(16, ceil(N_sample / 50)), N_sample is the total number of pixels in the labeled samples, and W and H are the width and height of the image, respectively; The specific steps of the category-balanced hierarchical allocation algorithm are as follows: S1-1. Count the number of labeled samples and the number of categories covered in each spatial block, and sort them from largest to smallest by the product of the number of samples and the number of categories; S1-2. Allocate spatial blocks to the training set, validation set, and test set in a 6:2:2 ratio, dynamically adjusting the proportion of each category in each set during the allocation process to ensure that the deviation from the global category proportion is less than 5%. S1-3. Ensure that there is at least one block buffer between the training block and the validation / test block to avoid overestimation of the evaluation caused by spatial autocorrelation.

3. The method according to claim 1, characterized in that, The criterion for identifying the minority class set in step S2 is as follows: C_minor = {c : n_c < α·max(n_c)} Where n_c is the number of samples in class c, α is the imbalance ratio threshold, adaptively set to 0.3; the number of enhancements to be synthesized for each class is aug_c = β·(max(n_c) - n_c), where β is the adaptive enhancement ratio, set to 0.8; The three-layer physical constraint verification specifically includes: S2-1, Value Range Constraint: The reflectance values ​​of all bands must be within the dynamic range of the sensor; if this is violated, the synthesized sample shall be discarded. S2-2, Smoothness constraint: If the sum of the absolute values ​​of the first-order differences of the synthetic sample spectral curves is less than a preset multiple of the average first-order difference of the original samples of the same type, and the sample is violated, it will be re-examined after spectral smoothing using the Savitzky-Golay smoothing kernel. If the sample is still violated, it will be discarded. S2-3. Diagnostic absorption peak constraint: The synthesized sample must have absorption characteristics that conform to the priori within the band range specified in the prior spectral knowledge base of the corresponding land cover category; otherwise, it shall be discarded. If the effective synthesis rate of a certain type of sample is less than 30% after three-layer constraints, then Gaussian perturbation enhancement of the spectral feature space is used instead of linear interpolation, with the formula x' = x_i + N(0, σ_c²·I), where σ_c is 0.05 times the standard deviation of each band of the c-th type of sample. After Gaussian perturbation, the three-layer physical constraints are also tested.

4. The method according to claim 1, characterized in that, The formula for the comprehensive information content score of each band in step S3 is as follows: Score(b) = w1·I(b) + w2·E(b) + w3·RSI(b) Where I(b) is the information entropy of band b, E(b) is the redundancy index between band b and adjacent bands, RSI(b) is the scene-specific relative spectral index contribution, and w1, w2, and w3 are the corresponding weights, and the sum of the three is 1. The steps for automatically determining the application scenario type include: statistically analyzing the average first derivative values ​​of various average spectra at the red edge position in the training samples; if the absolute value is greater than the first threshold, it is determined to be a vegetation-dominated scenario; statistically analyzing the absorption depth of the short-wave infrared absorption feature region; if it is greater than the second threshold, it is determined to be a mineral-dominated scenario; statistically analyzing the reflectance ratio of the visible light region to the near-infrared region; if it meets the water body reflectance characteristics, it is determined to be a water quality-dominated scenario; if none of the above are met, it is determined to be a general scenario. The method for adaptively determining the optimal number of bands K is as follows: sort all bands in descending order of their comprehensive information content scores, gradually accumulate the scores, and find the minimum K value that satisfies the condition that the proportion of the accumulated score to the total score is greater than or equal to a preset threshold T. At the same time, K is constrained to be within the upper and lower bounds of [max(10, floor(B / 10)), min(80, floor(B / 2))], where B is the original total number of bands. The threshold T is adaptively adjusted according to the average number of samples in each class, with T set to 0.90 for small sample scenarios and 0.95 for large sample scenarios.

5. The method according to claim 1, characterized in that, The calculation method for the four-dimensional comprehensive difficulty score in step S4 is as follows: D = 0.3·D1 + 0.3·D2 + 0.2·D3 + 0.2·D4 Wherein, D1 is the difficulty based on the number of categories, which is obtained by normalizing the number of categories; D2 is the difficulty based on the imbalance, which is calculated based on the Gini coefficient of the number of categories; D3 is the difficulty based on the spectral noise, which is calculated based on the mean square error before and after spectral smoothing; D4 is the difficulty based on the sample size, which is calculated based on the average number of samples per class; the difficulty score D ranges from [0,1]. The rules for matching candidate model structures in the predefined model pool based on difficulty level are as follows: when 0 ≤ D < 0.3, small-capacity models are matched, including three-layer one-dimensional convolutional neural networks and radial basis kernel support vector machines; when 0.3 ≤ D < 0.6, medium-capacity models are matched, including four-layer three-dimensional convolutional neural networks and hybrid convolutional neural networks; when 0.6 ≤ D ≤ 1.0, large-capacity models are matched, including lightweight spectral Transformer and spatial spectral fusion Transformer. The search space for Bayesian optimization includes the learning rate, weight decay coefficient, batch size, optimizer type, and data augmentation probability. The number of trials is adaptively adjusted according to the difficulty score D, with 40 trials for high difficulty and 12 trials for low difficulty. Each trial run performs only 15 rounds of rapid validation.

6. The method according to claim 1, characterized in that, The specific conditions for the three types of parallel detection in step S5 are as follows: S5-1, Early Stopping Detection of Normal Convergence: If the Kappa coefficient of the validation set does not increase by more than the increase threshold for a consecutive preset number of rounds, early stopping is triggered, and the model weights of the optimal round are saved. S5-2 Training Oscillation Detection: If the Kappa coefficient of the validation set fluctuates within a predetermined range for a continuous preset number of rounds and the current Kappa is lower than 0.9 times the historical baseline Kappa of the same difficulty, the learning rate will be automatically reduced to 1 / 3 of the original value, and training will continue from the current optimal weight. The learning rate will be reduced a maximum of two times. S5-3 Training divergence detection: If the training loss is greater than the divergence detection threshold, or the validation set Kappa is lower than the minimum qualified threshold and there is no upward trend in the previous few rounds, then the current training is terminated immediately, the random seed is replaced and the training is restarted using the suboptimal hyperparameter combination. The maximum number of training rounds, early stopping threshold, oscillation detection threshold, divergence detection threshold, and maximum number of automatic retries are all adaptively set in three levels according to the comprehensive difficulty score D: a stricter threshold and fewer retries are set when D < 0.3 in simple scenarios, and a more lenient threshold and more retries are set when D ≥ 0.6 in difficult scenarios. It also includes an optional cross-domain adaptive fine-tuning step S0, which is performed on unlabeled images in the target domain after training a pre-trained model with labeled images in the source domain. The first stage is statistical alignment based on spectral physical consistency, with the alignment formula being H'_{t,b} = ((H_{t,b} - μ_{t,b}) / σ_{t,b})·σ_{s,b} + μ_{s,b}, where μ_{s,b} and σ_{s,b} are the mean and standard deviation of band b in the source domain, and μ_{t,b} and σ_{t,b} are the mean and standard deviation of band b in the target domain. The second stage is semi-supervised fine-tuning based on consistency regularization, freezing the first 70% of the layers of the pre-trained model and fine-tuning only the last 30% of the top-level classifier. The total loss is L = L_CE + λ·L_consist, where the consistency loss weight λ is 0.3 across time phases and 0.8 across sensors.

7. An adaptive training system for a land cover classification model based on hyperspectral imagery, characterized in that, Includes the following modules: The adaptive spatial block partitioning module is used to read hyperspectral images and user-annotated sample sets. It adaptively calculates the block side length according to the image size and the number of samples, and divides the image into non-overlapping spatial blocks. The class-balanced hierarchical allocation algorithm is used to allocate the spatial blocks to the training set, validation set and test set in a 6:2:2 ratio, ensuring that the training block and the validation / test block are completely non-adjacent in space. The physically constrained imbalanced sample adaptive augmentation module is used to count the number of samples of each class in the training set, identify the minority class set, perform synthetic augmentation on the minority class samples, and apply three layers of physical constraint checks: value range constraint, smoothness constraint, and diagnostic absorption peak constraint. Only synthetic samples that pass all the checks are retained. When the effective synthesis rate is less than 30%, it automatically switches to the Gaussian perturbation augmentation scheme and outputs a class-balanced augmented training set. The scene-aware adaptive band selection module is used to calculate the comprehensive information content score of each band, automatically determine the application scene type and adjust the weight of the scene-specific spectral index accordingly, and adaptively determine the optimal number of bands and band index set based on the information content accumulation ratio and sample size. The dataset difficulty-driven model structure and hyperparameter adaptive search module is used to calculate the four-dimensional comprehensive difficulty score of the augmented training set, including the difficulty of class number, imbalance difficulty, spectral noise difficulty, and sample size difficulty. Based on the difficulty level, it matches candidate model structures in a predefined model pool and automatically searches for the optimal training hyperparameters under the Bayesian optimization framework. The training scheduling module with dynamic early stopping and automatic retry is used to adaptively set the maximum number of training rounds, early stopping threshold, oscillation detection threshold and divergence detection threshold according to the comprehensive difficulty score. After each round of training, three types of detection are performed in parallel: normal convergence early stopping detection, training oscillation learning rate reduction recovery detection and training divergence automatic retry detection, and output the finally trained ground cover classification model. The unified scheduling interface module is used to schedule the execution of the above modules in sequence and receive feedback signals from each module to adaptively adjust parameters, including adjusting the upper and lower bounds of the block size based on the uniformity of spatial block partitioning, adjusting the enhancement ratio and constraint threshold based on the enhancement efficiency, adjusting the number of retained bands based on the band efficiency, mapping the model pool and search space based on the difficulty score, adjusting the consistency loss weight based on the cross-domain loss, and deciding whether to stop early, reduce the learning rate, or retry based on the training status.

8. The system according to claim 7, characterized in that: In the adaptive spatial block partitioning module, the block side length formula is s = max(32, floor(sqrt(W×H / K_block))), where K_block = max(16, ceil(N_sample / 50)). The hierarchical allocation algorithm sorts the samples by the product of the number of samples and the number of categories and then allocates them in a 6:2:2 ratio to ensure a buffer interval of at least one block between the training block and the validation / test block. In the physical constraint imbalance sample adaptive enhancement module, the minority class judgment condition is n_c < 0.3·max(n_c), the enhancement quantity is aug_c = 0.8·(max(n_c) - n_c), the three-layer test includes value range test, smoothness test and diagnostic absorption peak test, and when the effective synthesis rate is less than 30%, the Gaussian perturbation scheme x' = x_i + N(0, σ_c²·I) is switched. The scene-aware adaptive band selection module uses a score of Score(b) = w1·I(b) + w2·E(b) + w3·RSI(b) to automatically determine vegetation, minerals, water quality or general scene type through red edge features, short-wave infrared absorption depth and visible light / near-infrared reflectance ratio. The optimal number of bands K is adaptively determined by the information accumulation ratio threshold T and the upper and lower bounds of the number of bands. In the model search module, the difficulty score D = 0.3·D1 + 0.3·D2 + 0.2·D3 + 0.2·D4 is divided into three levels: easy (D<0.3), medium (0.3≤D<0.6), and difficult (D≥0.6), which are matched with small, medium, and large capacity model pools respectively. The number of Bayesian optimization trials is adaptively adjusted according to the difficulty. The training scheduling module performs three types of detection: early stop, oscillating learning rate reduction, and divergent retry. The threshold is adaptively graded, and the random seed is changed and the suboptimal hyperparameter combination is switched during retry. The system also includes a cross-domain adaptive fine-tuning module, which is used to perform spectral physical consistency statistical alignment and semi-supervised consistency fine-tuning on the unlabeled images in the target domain based on the source domain pre-trained model, freeze 70% of the layers and fine-tune 30% of the layers, and adaptively select the consistency loss weight λ according to the cross-domain type.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the adaptive training method for land cover classification models based on hyperspectral imagery as described in any one of claims 1 to 6.

10. An electronic device comprising a processor and a memory, the memory for storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the adaptive training method for land cover classification model based on hyperspectral imagery as described in any one of claims 1 to 6.

Images