A universal crop mapping loss function based on phenology prior

Abstract

The application discloses a general crop mapping loss function based on phenology prior, which comprises crop phenology prior cross-entropy loss, and is verified by time series mapping of rape. The crop phenology prior cross-entropy loss comprises two key steps: firstly, the phenological characteristics of rape are extracted through a comprehensive vegetation index; and then, the loss is optimized by using the phenological characteristics as prior knowledge, so as to pay more attention to the identification of rape pixels. The overall framework of the rape mapping algorithm mainly comprises two parts: prior supervision and a deep learning model; the deep learning model comprises an LSTM, a 1DCNN and a DNN. The general crop mapping loss function based on the phenology prior is simple in technology and easy to operate; it provides a new solution for improving the precision of crop mapping of a deep learning model without changing the network structure and increasing the calculation amount; in addition, the idea of crop mapping by using the phenology prior is universal.

Description

TECHNICAL FIELD

[0001] The application relates to a loss function for crop mapping, in particular to a general crop mapping loss function based on phenology prior, and belongs to the technical field of deep learning crop mapping. BACKGROUND

[0002] Accurate and timely acquisition of crop distribution information is a basic prerequisite for accurate estimation and prediction of yield, and is of great significance to agricultural production, farmland management and national food security. As we all know, remote sensing technology is a low-cost, timely and efficient means of large-area crop monitoring, and has therefore received extensive attention in the past few decades. Existing research shows that making full use of the temporal information of remote sensing images can improve the performance of crop mapping. At present, many satellite remote sensing-based methods have been developed using multi-temporal images, mainly including two categories: vegetation index-based methods, such as normalized difference vegetation index and normalized difference vegetation index; and machine learning-based methods, such as random forest and support vector machine. These methods are simple to operate and easy to use, but their performance is limited by the artificially selected features, and they can only be used for rough feature representation, making it difficult to achieve high-precision crop mapping.

[0003] In recent years, with the rapid development of deep learning technology, great progress has been made in the research of time series image crop mapping. In terms of processing time series data, convolutional neural network (CNN), deep neural network (DNN) and long short-term memory network (LSTM) are all representative models. In order to further improve the accuracy, current research mainly considers introducing many complex structures, such as attention mechanism, combination of CNN and LSTM, etc. These methods improve the network performance by designing different advanced network structures [5], but at the same time lead to network redundancy and increase the computational complexity; especially for long time series images, the amount of calculation will increase significantly.

[0004] Besides improving network structure, loss function also plays an important role in deep learning. For neural networks that need supervision, a proper loss function can guide the network to train in the right direction, thus improving the performance of the network, so the design of new loss function is attracting more and more attention. Cross-entropy loss is the most widely used loss function at present, which can correctly classify common samples, but it is powerless for rare samples. In order to solve this problem, Li et al. proposed a hybrid loss combining weighted binary cross-entropy loss and dice coefficient loss to focus on the learning of rare samples [7]. Lin et al. proposed Focal loss to reduce the weight of loss assigned to easy samples to solve the problem of extreme imbalance of classification. In its technical solution, the loss first introduces a weight factor to balance the importance of positive and negative samples, and then introduces a modulation factor to reduce the loss contribution from easy samples and expand the range of sample calculation to obtain the loss. Specifically, for samples with small loss, compared with cross-entropy loss, the corresponding loss will be smaller after the loss function, so as to focus on the learning of difficult samples. The above improved loss function has better performance, but they mainly focus on solving the problem of sample imbalance, that is, the loss optimization of general deep learning tasks. Crops have a stable growth cycle and are planted and harvested regularly by humans. Remote sensing time series data can effectively monitor these phenological features, which is of great guiding significance for crop identification. However, the current loss function ignores the phenological differences between different land types. Many studies have shown that phenological information is an important feature to distinguish different types of crops, and adding prior knowledge to the loss can reduce the number of positive samples that are incorrectly identified. However, fully exploiting phenological information as prior to optimize the loss function and improve the performance of the model in the crop mapping task of time series images has not been fully explored.

[0005] (1) The existing deep learning model can improve the performance of the model by designing advanced modules, but it inevitably leads to an increase in complex calculation and network redundancy, especially when long time series images are used.

[0006] (2) For methods based on loss function to improve performance, most existing loss functions generally focus on loss optimization, that is, the problem of sample imbalance, ignoring the particularity of the task.

[0007] (3) The existing multi-source remote sensing image fusion algorithm is complex, which limits it to only use limited data sources. SUMMARY

[0008] The purpose of the present application is to provide a general crop mapping loss function based on phenological prior to solve the problems raised in the background art.

[0009] To achieve the above object, the present application provides the following technical scheme: a general crop mapping loss function based on phenological prior, including crop phenological prior cross entropy loss (PPCE), and verifying the time series mapping of rape;

[0010] The crop phenological prior cross entropy loss includes two key steps: first, extracting the phenological characteristics of rape through the comprehensive vegetation index, and then using the phenological characteristics as prior knowledge for loss optimization to pay more attention to the identification of rape pixels;

[0011] Among them, the overall framework of the rape mapping algorithm mainly includes two parts: prior supervision and deep learning model;

[0012] The deep learning model includes LSTM, 1DCNN and DNN.

[0013] As a preferred technical scheme of the present application, the prior supervision includes the following steps:

[0014] Step 1: Extraction of phenological prior knowledge

[0015] The phenological characteristics of rape mainly include two parts:

[0016] (1) In the flowering period, the rape flowers are yellow, and the corresponding pixels on the remote sensing image also show yellow;

[0017] (2) In other growth periods, the chlorophyll in the rape leaves has a strong reflection effect on green light, so it appears green in the image;

[0018] As prior knowledge, these phenological characteristics are helpful to distinguish rape pixels from other pixels, so they are used in the present application. Vegetation index has been widely used in the representation of crop phenological characteristics. In order to capture the phenological characteristics of rape in the flowering period, the present application selects normalized difference yellow index (NDYI) to extract the prominent yellow tone, and uses normalized vegetation index (NDVI) to ensure that the leaves of rape are green in other growth periods. NDYI is positively correlated with the carotenoid content in rape flowers, and NDVI is positively correlated with the chlorophyll content in leaves; therefore, NDYI and NDVI can quantitatively describe the phenological characteristics of the corresponding pixels. The higher the NDYI value, the higher the probability of yellow flowers, and similarly, the higher the NDVI value, the higher the probability of green leaves in other growth periods.

[0019]

[0020] Where G and B represent the reflectance of green and blue bands at flowering stage; R, NIR represent the reflectance of red and near-infrared bands at other growth stages, NDYI, NDVI range from -1 to 1.

[0021] Therefore, the combination of NDYI and NDVI can obtain a relatively reliable estimate of the probability of rape, that is, the joint probability of a pixel is the comprehensive reflection of the rape phenology characteristics in the corresponding "yellow" and "green" characteristics at flowering and other growth stages. Specifically, for a given pixel, the multiplication of the corresponding NDYI and NDVI will obtain a rape probability distribution map. In order to avoid the situation of positive negative multiplication when both vegetation indices are negative, the range of NDYI and NDVI is converted from [-1, 1] to [0, 1] by using the normalization method. In addition, in order to increase the difference between rape pixels and other pixels, the Sigmoid function is used for nonlinear conversion and normalization operation on the above obtained probability map to obtain the final rape probability map.

[0022] Second step: PPCE loss construction

[0023] Using appropriate loss function for model training can help improve the prediction accuracy of the model. Cross-entropy loss is the most commonly used loss function in deep learning crop mapping tasks.

[0024] Using appropriate loss function for model training can help improve the prediction accuracy of the model. Cross-entropy loss is the most commonly used loss function in deep learning crop mapping tasks.

[0025] Its definition is as follows:

[0026]

[0027] Where N represents the number of training samples, y i represents the one-hot encoding of the ith label, p i represents the predicted probability value of the model for the ith sample.

[0028] Cross-entropy loss can achieve satisfactory results under the condition of sample balance. However, when there is a serious sample imbalance problem, the cross-entropy loss function may be ineffective. Since there is a serious imbalance between rape pixels and other pixels in this study, it is not suitable for this task. In order to solve this problem, the invention introduces a weight factor for each class to adjust the proportion of rape and other samples. The modified loss function is defined as k-balanceCE (KCE), which can be represented as:

[0029]

[0030] Wherein k1 represents the weight of the rape sample, k2 represents the weight of other categories, and k1+k2=1.

[0031] Although the introduction of weight factors k1 and k2 can alleviate the sample imbalance problem, it ignores the difference between the prediction probability value of the misclassified sample and the true label. Crop phenology prior is beneficial to the correct identification of crops. Therefore, in this paper, the crop phenology information of each sample is used as a priori in the loss function to enhance this difference. It mainly targets the rape pixels and increases their contribution to the final loss, thereby enhancing the network's ability to identify rape pixels. The PPCE constructed can be expressed as:

[0032]

[0033] Wherein W represents the weight matrix of the rape probability generation, and the value range is 0-1. From the above formula, it can be seen that in order to achieve a small loss, when a pixel corresponds to a larger W, the network will tend to predict the pixel as a rape pixel. If a rape pixel is misclassified into other categories, due to the larger rape probability obtained by the phenology prior, the contribution to the final loss will be larger, thereby guiding the network to pay more attention to the case of misidentification of rape pixels.

[0034] As a preferred technical solution of the present application, the deep learning model is used to fit the complex correlation between the input features and the true labels. Through gradient descent algorithm iterative optimization of model weight, the error between the predicted value and the true value is minimized. The loss function is used to represent the difference between the model predicted value and the true value. By reducing the loss function, the model parameters can be optimized.

[0035] As a preferred technical solution of the present application, the LSTM is a variant of recurrent neural network (RNN) model, which is specially used for the expression of sequence data. LSTM is mainly composed of forget gate, input gate and output gate, which control the cell state and output.

[0036] As a preferred technical solution of the present application, the 1DCNN has been proved to be effective in processing the time dimension in time series classification tasks, by stacking multiple convolution layers, the shallow layers of the network can capture local features, and the deep layers of the network can obtain abstract features on a larger scale; by fusing features of different scales, the discriminability of the features can be enhanced, thereby enhancing the performance of the network; the 1DCNN mainly consists of convolution layers, pooling layers, fully connected layers and output layers; the specific implementation process is to first input the time series data as the input of the input layer, and then use the convolution layers and the pooling layers to perform feature extraction and sparse processing layer by layer, wherein each convolution layer is composed of multiple convolution kernels, the convolution kernels in the same layer are of the same size, the pooling layer adopts the average pooling method, and then the results are classified by the fully connected layer.

[0037] As a preferred technical solution of the present application, the DNN is an artificial neural network (ANN), which has no fixed structure and is fully connected between layers, that is, the neurons in a layer are fully connected with all the neurons in the adjacent layer, which can be used as a basic model for evaluating the classification performance of time series. In addition to the input layer and the output layer, the DNN also includes one or more hidden layers.

[0038] Compared with the prior art, the present application has the following advantages: the present application provides a general crop mapping loss function based on phenology prior, which is different from the prior art. The prior art generally focuses on loss optimization, that is, the problem of sample imbalance. For example, the particularity of the crop mapping task is ignored, that is, crops generally have unique phenological characteristics that can be used as prior knowledge to guide model training. The present application provides a general crop mapping loss function based on phenology prior. This method has the following advantages:

[0039] (1) The method is simple to use and easy to operate. It provides a new solution for improving the accuracy of crop mapping for deep learning models without changing the network structure or increasing the computational load.

[0040] (2) The idea of using phenology prior for crop mapping is universal. Because crops often show different phenological priors, it can be easily extended to other mapping tasks.

[0041] (3) The function provides a novel way of integrating multi-source remote sensing images. It can fully utilize the advantages of high spatial resolution and large coverage of multi-source remote sensing images to improve the mapping ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 The present application provides a general crop mapping loss function optimization method. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application are clearly and completely described, obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0044] Please refer to Figure 1 The present application provides a kind of general crop mapping loss function based on phenology prior, including crop phenology prior cross entropy loss (PPCE), and with rape in time series mapping is verified;

[0045] Crop phenology prior cross entropy loss includes two key steps: first, the phenological characteristics of rape are extracted by comprehensive vegetation index, then, the loss optimization is carried out using the phenological characteristics as prior knowledge, to pay more attention to the identification of rape pixels;

[0046] Among them, the overall framework of rape mapping algorithm mainly includes two parts: prior supervision (prior supervision) and deep learning model (deep learning model);

[0047] The deep learning model (deep learning model) includes LSTM, 1DCNN and DNN;

[0048] The prior supervision includes the following steps:

[0049] First step: extraction of phenological prior knowledge

[0050] The phenological characteristics of rape mainly include two parts:

[0051] (1) In the flowering period, rape flowers are yellow, and the corresponding pixels on the remote sensing image also show yellow;

[0052] (2) In other growth periods, chlorophyll in rape leaves has strong reflection to green light, so it shows green in the image;

[0053] As prior knowledge, these phenological features help to distinguish the pixels of oilseed rape from other pixels, and thus are utilized in the present application. Vegetation indices have been widely used for the characterization of crop phenological features. To capture the phenological features of oilseed rape at the flowering stage, the present application selects the normalized difference yellow index (NDYI) to extract the prominent yellow hue, while using the normalized vegetation index (NDVI) to ensure that the leaves of oilseed rape are green at other growth stages. NDYI is positively correlated with the carotenoid content in the flowers of oilseed rape, and NDVI is positively correlated with the chlorophyll content in the leaves; therefore, NDYI and NDVI can quantitatively describe the phenological characteristics of the corresponding pixels. The higher the NDYI value, the higher the probability of yellow flowers. Similarly, the higher the NDVI value, the higher the probability of green leaves at other growth stages.

[0054]

[0055] where G and B represent the reflectance of the green and blue bands at the flowering stage; R, NIR represent the reflectance of the red and near-infrared bands at other growth stages. The NDYI, NDVI value range is -1~1.

[0056] In view of this, the combination of NDYI and NDVI can obtain a relatively reliable estimate of the probability of oilseed rape, i.e. the joint probability of a pixel is the comprehensive reflection of the corresponding "yellow" and "green" features of oilseed rape phenological features at the flowering stage and other growth stages. Specifically, for a given pixel, multiplying the corresponding NDYI and NDVI will obtain an oilseed rape probability distribution map. In order to avoid the case of positive when two vegetation indices are negative, the method of normalization is used to convert the range of NDYI and NDVI from [-1, 1] to [0, 1]. In addition, in order to increase the difference between oilseed rape pixels and other pixels, the Sigmoid function is used for nonlinear transformation and normalization operation on the above-mentioned probability map to obtain the final oilseed rape probability map. It can be expressed as follows:

[0057] W = Norm (Sig (NormNDYI) * Norm (NDVI)) (3)

[0058] where Sig represents the Sigmoid function, Norm represents the normalization operation, and W represents the extracted crop phenological prior

[0059] Second step: PPCE loss construction

[0060] Using appropriate loss functions for model training can help improve the prediction accuracy of the model. Cross-entropy loss is the most commonly used loss function in deep learning crop mapping tasks.

[0061] which is defined as follows:

[0062]

[0063] where N represents the number of training samples, y i represents the one-hot encoding of the ith label, p i represents the predicted probability value of the model for the ith sample. t

[0064] The cross-entropy loss can achieve satisfactory results in the case of sample balance. However, when there is a serious sample imbalance problem, the cross-entropy loss function may be ineffective. Since there is a serious imbalance between the number of rape pixels and other pixels in this study, it is not suitable for this task. In order to solve this problem, the invention introduces a weight factor for each class to adjust the proportion of rape and other samples. The modified loss function is defined as k-balanceCE (KCE), which can be represented as:

[0065]

[0066] where k1 represents the weight of rape samples, k2 represents the weight of other classes, and k1+k2=1.

[0067] The introduction of weight factors k1 and k2 can alleviate the sample imbalance problem, but it ignores the difference between the predicted probability value of the misclassified sample and the true label. Crop phenology prior is beneficial to the correct identification of crops. Therefore, this paper takes the crop phenology information of each sample as a prior in the loss function to enhance this difference. It mainly targets rape pixels and increases their contribution to the final loss, thereby enhancing the network's ability to identify rape pixels. The constructed PPCE can be represented as:

[0068]

[0069] where W represents the weight matrix of rape probability generation, and its value range is 0-1. As can be seen from the above formula, in order to achieve a smaller loss, when a pixel corresponds to a larger W, the network will tend to predict that pixel as a rape pixel. If a rape pixel is misclassified into other classes, due to the larger rape probability obtained by the phenology prior, the contribution to the final loss will be larger, thereby guiding the network to pay more attention to the case of misidentification of rape pixels.

[0070] The deep learning model is used to fit the complex correlation between the input features and the true labels, and the model weights are iteratively optimized through the gradient descent algorithm, so that the error between the predicted value and the true value is minimized. The loss function is used to represent the difference between the predicted value and the true value of the model, and by reducing the loss function, the model parameters can be optimized.​

[0071] The LSTM is a variant of recurrent neural network (RNN) model, which is specially used for the representation of sequence data. The LSTM is mainly composed of forget gate, input gate and output gate, which control the cell state and output;

[0072] The detailed formulas are as follows:

[0073] f t = σ (W f · [H t-1 , x t ] + b f ) (7)

[0074] c t = tanh (W c · [H t-1 , x t ] + b c ) (8)

[0075] i t = σ (W i · [C t-1 , x t ] + b i ) (9)

[0076] o t = σ (W o · [H t-1 , x t ] + b o ) (10)

[0077] H t = o t · tanh (c t ) (11)

[0078] Where c t is the cell state value. The cell state is used to save the accumulation of valid information from time step 1 to time step t, and update and utilize the information under the control of the three gates. The symbol ⊙ represents matrix dot multiplication. tanh represents the hyperbolic tangent function. i t is the input gate, which is used to control how much information is obtained from the current time step t. W i and b i are the weight and offset vector of the input gate, respectively. C t-1 and H t-1 are the cell state and hidden state at time t-1. x t represents the deep feature obtained at step t-1. σ represents the sigmoid activation function. The forget gate f tDecides how much information to keep from the previous time step. Where W f and b f are the weight and bias vector of the forget gate, respectively. t Decides what information the LSTM will output. Where W o and b o are the weight and bias vector of the output gate, respectively. H t represents the value of the hidden layer.

[0079] Wherein the 1DCNN has been proved to be effective in processing the time dimension in the time series classification task, by stacking multiple convolutional layers, the shallow layers of the network can capture local features, and the deep layers of the network can obtain abstract features on a larger scale; by fusing features of different scales, the discriminability of the features can be enhanced, thereby enhancing the performance of the network; the 1DCNN mainly consists of convolutional layers, pooling layers, fully connected layers, and output layers; the specific implementation process is to first input the time series data as the input of the input layer, and then use the convolutional layers and the pooling layers to perform feature extraction and sparse processing layer by layer, wherein each convolutional layer consists of multiple convolutional kernels, the convolutional kernels in the same layer are of the same size, the pooling layer adopts the average pooling method, and then the results are classified by the fully connected layer

[0080]

[0081] Wherein k represents the convolution kernel, j represents the number of convolution kernels, M represents the number of input features, b represents the bias of the corresponding convolution kernel, f() represents the activation function, and * represents the convolution operation. Further, the DNN is an artificial neural network (ANN) which has no fixed structure, and the layers are fully connected, i.e. the neurons in a layer are fully connected to all neurons in the adjacent layer, which can be used as a basic model for evaluating the classification performance of time series. In addition to the input layer and the output layer, the DNN also includes one or more hidden layers, and the function of the hidden layer is to convert the original information into meaningful features for classification. A complete DNN can be represented as:

[0082] f(x)=G(b (i) +W (i) …(s(b (1) +W (1) x))) (13)

[0083] Wherein x represents an n-dimensional vector, and n depends on the dimension of the input data. W (i) represents the weight of the i-th hidden layer, and b (i) represents the bias of the i-th hidden layer. s represents the activation function, commonly used are the relu function or the tanh function; G represents the Sigmoid function.

[0084] While embodiments of the present application have been shown and described, it is to be understood that the embodiments described are merely divergences of the principles and spirit of the present application and that numerous modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A general crop mapping loss function optimization method based on phenological priors, characterized in that, This includes the crop phenological prior cross-entropy loss PPCE, which is validated using rapeseed in time series mapping; The crop phenological prior cross-entropy loss includes two key steps: First, extract the phenological features of rapeseed by integrating vegetation indices. Then, use the phenological features as prior knowledge to optimize the loss, so as to pay more attention to the recognition of rapeseed pixels. The overall framework of the rapeseed mapping algorithm consists of two main parts: prior supervision and deep learning model. The deep learning models include LSTM, 1DCNN, and DNN; Among them, PPCE loss construction Using a loss function for model training helps improve the model's prediction accuracy. Cross-entropy loss is the most commonly used loss function in deep learning crop mapping tasks. Its definition is as follows: Where N represents the number of training samples, y i The one-hot encoding of the ith tag, p i This represents the predicted probability value of the model for the ith sample; Cross-entropy loss can achieve satisfactory results when the sample is balanced. However, it may fail when there is a severe imbalance in the number of rapeseed pixels and other pixels. A weighting factor is introduced for each class to adjust the ratio of rapeseed to other samples. The modified loss function is defined as KCE, which can be expressed as: Where k1 represents the weight of the rapeseed sample, k2 represents the weight of other categories, and k1+k2=1; Introducing weighting factors k1 and k2 can alleviate the sample imbalance problem, but it ignores the difference between the predicted probability value and the true label of misclassified samples. Crop phenological priors are beneficial to the correct identification of crops. Therefore, the crop phenological information of each sample is used as a prior in the loss function to enhance this difference. The crop phenological information is applied to rapeseed pixels and its contribution to the final loss is increased, thereby enhancing the network's ability to identify rapeseed pixels. The constructed PPCE can be expressed as: Where W represents the weight matrix for generating rapeseed probability, and its value ranges from 0 to 1. As can be seen from the above formula, in order to achieve a smaller loss, when the W corresponding to a pixel is large, the network will tend to predict that pixel as a rapeseed pixel. If a rapeseed pixel is misclassified into another class, the rapeseed probability obtained from the phenological prior will be large, and its contribution to the final loss will be large, thus guiding the network to pay more attention to the situation where rapeseed pixels are misidentified.

2. The method for optimizing a general crop mapping loss function based on phenological priors according to claim 1, characterized in that: The prior supervision includes the following steps: Step 1: Extraction of prior knowledge of phenology The phenological characteristics of rapeseed include two parts: (1) During the flowering period, rapeseed flowers are yellow, and the corresponding pixels in the remote sensing image are also yellow; (2) During other growth stages, the chlorophyll in rapeseed leaves has a strong reflective effect on green light, so it appears green in the image. As prior knowledge, these phenological characteristics help distinguish rapeseed pixels from other pixels. Vegetation indices have been widely used to characterize crop phenological features. To capture the phenological characteristics of rapeseed during its flowering period, the Normalized Yellowness Index (NDYI) was selected to extract prominent yellow hues. At the same time, the Normalized Vegetation Index (NDVI) was used to ensure that rapeseed leaves are green during other growth stages. NDYI is positively correlated with the carotenoid content in rapeseed flowers, and NDVI is positively correlated with the chlorophyll content in leaves. Therefore, NDYI and NDVI can quantitatively describe the phenological characteristics of the corresponding pixels. The higher the NDYI value, the higher the probability of yellow flowers. Similarly, the higher the NDVI value, the higher the probability of green leaves during other growth stages. Where G and B represent the reflectance of the green and blue bands during the flowering period; R and NIR represent the reflectance of the red and near-infrared bands during other growth stages; and NDYI and NDVI range from -1 to 1. Therefore, the combination of NDYI and NDVI can provide a relatively reliable estimate of the rapeseed probability. That is, the joint probability of a pixel is a combination of the phenological characteristics of rapeseed reflected in the corresponding "yellow" and "green" features during the flowering period and other growth stages. Specifically, for a given pixel, multiplying the corresponding NDYI and NDVI will yield a rapeseed probability distribution map. To avoid the situation where the product of two negative vegetation indices is negative, a normalization method is used to convert the range of NDYI and NDVI from [-1, 1] to [0, 1]. In addition, to increase the difference between rapeseed pixels and other pixels, the probability map obtained above is subjected to nonlinear transformation and normalization operations using the Sigmoid function to obtain the final rapeseed probability map.

3. The method for optimizing a general crop mapping loss function based on phenological priors according to claim 1, characterized in that: The deep learning model is used to fit the complex correlation between input features and the true labels. The model weights are iteratively optimized through the gradient descent algorithm to minimize the error between the predicted and true values. The loss function is used to represent the difference between the model's predicted and true values. The model parameters can be optimized by reducing the loss function.

4. The method for optimizing a general crop mapping loss function based on phenological priors according to claim 1, characterized in that: The LSTM is a variant of the recurrent neural network (RNN) model, specifically designed for representing sequential data. LSTM consists of a forget gate, an input gate, and an output gate, which control the cell state and output.

5. The method for optimizing a general crop mapping loss function based on phenological priors according to claim 1, characterized in that: The 1DCNN has been proven effective in handling the temporal dimension in time series classification tasks. By stacking multiple convolutional layers, the shallow layers of the network can capture local features, while the deeper layers acquire abstract features at a larger scale. By fusing features at different scales, the discriminative power of the features can be enhanced, thereby improving the network's performance. The 1DCNN consists of convolutional layers, pooling layers, fully connected layers, and an output layer. The specific implementation process is as follows: first, time series data is used as input to the input layer, and then convolutional and pooling layers are used to extract features and perform sparsity processing layer by layer. Each convolutional layer consists of multiple convolutional kernels, and the kernels in the same layer are of the same size. The pooling layer uses the average pooling method, and then the fully connected layer classifies the results.

6. The method for optimizing a general crop mapping loss function based on phenological priors according to claim 1, characterized in that: The DNN is an Artificial Neural Network (ANN). It has no fixed structure, and the layers are fully connected, meaning that neurons in one layer are fully connected to all neurons in the adjacent layers. It can serve as a basic model for evaluating the classification performance of time series data. In addition to the input layer and output layer, the DNN also contains one or more hidden layers.

Images