CN105654210A - Ensemble learning fishery forecasting method utilizing ocean remote sensing multi-environmental elements - Google Patents

Abstract

The invention relates to the field of remote sensing information fishery application, in particular to an ensemble learning fishery forecasting method utilizing ocean remote sensing multi-environmental elements. The method aims at the problem that an existing fishery forecasting model is prone to be caught in overfitting on sample data, and consequently the generalization ability of the forecasting model is reduced, an ensemble learning method is adopted, a plurality of decision-making trees of simple structures are adopted as meta learning machines, learning machine integration is carried out based on a boosting algorithm, and the ensemble learning fishery forecasting method utilizing the ocean remote sensing multi-environmental elements is constructed. Each simple meta learning machine only learns a subset of characteristic space, the weight of samples, forecast to be wrong, in trained sub-learning machines as samples of the subsequent meta learning machines can be improved in the model training process to guarantee the different degree of the meta learning machines, and the learning machines learn information of different characteristic space subsets. According to the method, the generalization error can be reduced while prediction precision is improved, and the position of a fishery is effectively, fast and accurately located.

Description

Integrated learning fishery forecasting method utilizing ocean remote sensing multi-environment elements

Technical Field

The invention relates to the field of remote sensing information fishery application, in particular to an integrated learning fishery forecasting method utilizing ocean remote sensing multi-environment elements.

Background

The marine environment and fishery resource information is an important information support for marine fishery resource utilization and protection, and the development of fishery information service by utilizing marine remote sensing becomes an important guarantee for marine fishery production and management planning. The fishing situation distribution of the marine fishing ground is related to various factors such as natural environment, human activities and the like, and the forecast of the fishing ground distribution is influenced by various factors such as the accuracy of fishery production data recording and the accuracy of remote sensing inversion of environmental elements. The existing fishery forecasting method is to train a single learner model to establish a marine fishery forecasting model by utilizing a single statistical learning model aiming at a given training data set consisting of a few marine environment parameters such as sea surface temperature and fishery production data. From the aspect of a forecasting method, a traditional single statistical learning model needle belongs to a strong learning machine, and is easy to fall into overfitting sample data to reduce the generalization capability of a forecasting model. The mapping relation is established by adopting a single traditional machine learning model, and the problem of over-learning often exists, namely, the over-fitting of the mapping model to sample data causes that the fitting error of the mapping model to the sample is very small, but the generalization popularization performance is insufficient, and the error rate of actual data prediction which does not participate in learning training is very high.

The present invention incorporates the following publications:

[1] treble, wuyumei, zhangjing, zhou xing fang, fan wei, dong nan pacific pinus trachurus fishery forecast [ J ] based on classification regression tree algorithm, china oceanic university newspaper (nature science edition), 2012, Z2:53-59.

[2] Zhou Zhifeng, Xiang Wei, Cuxusen, Yanggonglong, Wuyumei, the Indian ocean major eye tuna fishery forecast based on Bayesian probability [ J ] fishery information and strategy, 2012,03: 214-.

[3] The method comprises the steps of (1) building a forecasting model of a Napexie Pacific ocean soft fish farm based on naive Bayes [ J ]. the academic newspaper of China ocean university (Nature science edition), 2015,02:37-43.

[4] Chenxue faithful, Fan Wei, trexuen, Zhou Feng, Tang Peak Hua, forecast [ J ] oceanic bulletin (Chinese version) based on Indian ocean Changfin tuna fishery in random forests, 2013,01: 158-.

Wherein [1] relates to a classification regression tree method, [2] and [3] relate to a Bayesian method, and [4] relates to a random forest method.

Disclosure of Invention

The invention aims to provide a learning fishery forecasting method based on integration of a plurality of sub-learning machines by utilizing ocean remote sensing multi-environment elements so as to improve generalization capability and forecasting precision of a fishery forecasting model.

In the present invention, the chinese meaning of CPUE is: the unit fishing effort is the fishing amount, and the English is cathperunit effort (CPUE).

The boosting algorithm is also a machine learning algorithm, is a method for improving the accuracy of the weak classification algorithm, and is applied to the field of computers. The boosting algorithm comes from:

YoavFreundandRobertE.Schapire.ADecision-TheoreticGeneralizationofOn-LineLearningandanApplicationtoBoosting,Journalofcomputerandsystemsciences,Volume55,Issue1,August1997,Pages119–139.

the invention relates to a fishery forecasting method based on ensemble learning, which comprises the following steps:

(1) the method comprises the steps of using sea surface environment data obtained by a satellite remote sensing technology as model input data to preprocess sea remote sensing multiple environment elements

And cutting the acquired global sea surface environment data obtained by satellite remote sensing according to the latitude and longitude range of the region of interest to obtain a region of interest environment data subset. The remote sensing inversion data is influenced by factors such as weather conditions, and numerical values are lost at different time periods and geographic positions. Missing values are interpolated based on the neighborhood correlation.

(2) Fishery fishing production data preprocessing

And (4) calculating the fishing amount of unit fishing effort according to the fishing production record data.

CPUE (yield of fisheries/effort of fishing)

And according to the statistical distribution of the CPUE and the fishery cost, establishing a classification label of the fishery by adopting a certain CPUE value as a threshold value of whether the CPUE value is the fishery. A record of the absence of fishing effort information in the fishery data is treated as an invalid sample (NA).

(3) Matching fishery production data with the marine remote sensing environment, establishing a multi-feature set of a constructed model, and searching marine environment data with the smallest space and time distance according to the time and longitude and latitude of the fishery production data to serve as the matching environment features of the sample.

(4) According to the layered sampling method, the samples are divided into ten parts, and the proportion of the number of the samples with classification labels of a fishing ground and a non-fishing ground in each part of the samples is kept basically consistent.

(5) A decision tree with a simple structure is used as a sub-learning machine, and learning machine integration is carried out based on a boosting algorithm.

The ensemble learning model includes M mutually independent sub-learning machines. The boosting integration method can be divided into the following three steps:

1) weight distribution W of initialized training data_m. For N learning samples, the weight of each training sample is initialized to W₁(i)＝1/N；

2) Sub-learning machine G for training M bases in sequence_m；

(i) Learning by using Wm as weight distribution of training data set to obtain basic classifier G_m(X), X → { -1, 1}, and calculating a classification error rate e thereof_mAnd classifier integration weight α_m：

$e_m=\underset{\left\{i:G_m\left(x_i\right)\≠y_i\right\}}{\mathrm{\Σ}}{w}_m\left(i\right)$

${\mathrm{\α}}_m=\frac{1}{2}\log \frac{1-e_m}{e_m}$

(ii) Updating weight distribution W of training data set_m+1:

$W_{m+1}\left(i\right)=\frac{W_m\left(i\right)}{Z_m}\exp (-{\mathrm{\α}}_m{y}_i{G}_m(x_i\left)\right)$

$Z_m=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{W}_m\left(i\right)\exp \left(-{\mathrm{\α}}_m{G}_{m\left(x_i\right)}{y}_i\right)$

In the training process, the distribution weight of the samples wrongly classified by the basic classifier Gm is increased, so that the samples are more easily selected by the subsequent sub-learning machines to serve as learning samples. The boosting integration method can focus on samples with difficult classification, meanwhile, the sample space subsets of the learning sub-sets are kept to have certain difference, and adaptive integration is carried out.

3) And carrying out weighted average on the results of the sub-learning machines to serve as a final integration model.

$G\left(x\right)=sign\left\{\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{\α}}_m{G}_m\left(x\right)\right\}$

The accuracy of each learning machine is 1-p, and the probability of errors of the integrated system is as follows:when p < 1/2, p_errThe integration precision is improved along with the increase of the learning machine number.

With V^α(x) The output of the α th learning machine under the input x is shown, and under the integration mode of weighted average, the output of the integration system is:the square of the error between the integrated output and the actual value is: $e\left(x\right)={\&lsqb;f\left(x\right)-\stackrel{\&OverBar;}{V}\left(x\right)\&rsqb;}^2.$ the degrees of deviation of the output of the learning machine α from the actual value squared error and from the integrated output are:^α(x)＝[f(x)-V^α(x)]² the total error of the individual output of the learning machine in the integrated system and the total deviation degree of the individual output of the learning machine in the integrated system and the integrated output are respectively as follows: $\begin{array}{cc}\mathrm{\&epsiv;}\left(x\right)={\mathrm{\&Sigma;}}_{\mathrm{\&alpha;}=1}^N{\mathrm{\&omega;}}_{\mathrm{\&alpha;}}{\mathrm{\&epsiv;}}^{\mathrm{\&alpha;}}\left(x\right)& a\left(x\right)={\mathrm{\&Sigma;}}_{\mathrm{\&alpha;}=1}^N{\mathrm{\&omega;}}_{\mathrm{\&alpha;}}{a}^{\mathrm{\&alpha;}}\left(x\right)\end{array}.$ under the input spatial density distribution p (x), the generalization error of the whole integrated system is: $E=\&Integral;p\left(x\right)\left(\mathrm{\&epsiv;}\right(x)-\mathrm{\&alpha;}(x\left)\right)dx=\stackrel{\&OverBar;}{E}-\stackrel{\&OverBar;}{A}.$ wherein,represents the average deviation of the system and represents the average deviation of the system,the degree of correlation of each learning machine is measured. The learning machines in the integration are independent from each other, the integration difference degree is large, and the generalization error of the integration is far smaller than the weighted average of the generalization errors of the learning machines.

The method adopts a plurality of decision trees with simple structures as sub-learning machines, carries out learning machine integration based on boosting algorithm, and constructs the fishing ground forecasting model based on ensemble learning. Each simple sub-learning machine learns only a subset of the feature space. In the model training process, the boosting algorithm can improve the weight of a sample with wrong prediction in the trained sub-learning machine as a sample of a subsequent sub-learning machine so as to ensure the difference degree of each sub-learning machine, and each learning machine learns different feature space subset information. Therefore, it is possible to reduce the generalization error while improving the prediction accuracy.

(6) Fishing ground forecasting using established models

And based on the constructed model, forecasting the fishing ground according to the input marine environment information.

Due to the adoption of the technical scheme, compared with the prior art, the invention has the following advantages and positive effects: the invention adopts a plurality of simple decision trees as the sub-learning machines to learn the input characteristics and samples, each sub-learning machine only learns partial characteristic subsets of the input, and integrates each sub-learning machine by adopting a boosting method to construct an integrated learning fishery forecasting method utilizing ocean remote sensing multiple environmental elements, thereby avoiding the over-fitting problem caused by a complex model, improving the forecasting performance and generalization and popularization capability of the forecasting model, effectively and quickly determining the position of the fishery, reducing the fish searching time of the fishery, improving the yield of fishery and having important scientific value, economic value and social benefit.

Drawings

FIG. 1 is a flow chart of a fishing ground forecasting method of multi-factor integrated learning according to the method of the present invention.

FIG. 2 shows spatial distribution of fishery samples according to an embodiment of the present invention.

FIG. 3 shows probability distributions and correlations of various feature data according to an embodiment of the present invention.

FIG. 4 shows the fishing ground probability prediction results according to the embodiment of the present invention.

Detailed Description

The invention will be further illustrated with reference to the following specific examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the present invention. Further, it should be understood that various changes or modifications of the present invention may be made by those skilled in the art after reading the teaching of the present invention, and such equivalents may fall within the scope of the present invention as defined in the appended claims.

The embodiment of the invention relates to a method for forecasting an integrated learning fishing ground by utilizing marine remote sensing multi-environment elements, wherein an interested area is 105-130E degrees and 0-30N longitude and latitude in the south China sea, and the method comprises the following steps:

(1) preprocessing multiple environmental elements of ocean remote sensing by taking the sea surface environmental data (chlorophyll a, sea surface temperature and sea surface height) obtained by the satellite remote sensing technology as model input data

Chlorophyll a (chlorea), Sea Surface Temperature (SST), sea surface height and other marine environmental parameter data. Wherein the chlorophyll a and sea surface temperature comprise the global inversion data of MODIS satellite aquaa and Terra spatial resolution 4km time resolution 8 days, and the sea surface temperature comprises products of SST, NSST, SST4 and other three different algorithms. Sea surface height is data for a south sea area with spatial resolution of 0.25 degrees temporal resolution of 1 day.

The original marine environment data is 2009-2014 MODIS satellite 4-kilometer resolution 8 balance global data, and comprises 2198 ERDASImagine format files and 2208 HDF format files in total.

Data cutting: and according to the conventional operation range of the south sea fishery, cutting according to the latitude and longitude ranges of 105E-130E and 0N-30N to obtain the environment data subset of the south sea operation area.

Data interpolation: the remote sensing inversion data is influenced by factors such as weather conditions, and numerical values are lost at different time periods and geographic positions. Missing values are interpolated based on the neighborhood correlation.

(2) Fishery fishing production data preprocessing

Collecting production data of iris tectorum to obtain 939 production records, and calculating the unit effort of catch according to the input catch yield and production duration

CPUE-yield/duration of catch production

According to the statistical distribution of the CPUE, the number of digits (mediavalue) is used as a threshold value for judging whether the CPUE is a fishery, and a fishery classification label is established.

The total number of records with missing production time information in the fishery data is 50, and the records are treated as invalid samples (NA).

(3) Matching fishery production data with a marine remote sensing environment,

and searching the marine environment data with the minimum space and time distance according to the operation date and latitude and longitude of the fishery production data, and taking the marine environment data as the matching environment characteristics of the sample.

(4) According to the layered sampling method, the samples are divided into ten parts, and the proportion of the number of the samples with classification labels of a fishing ground and a non-fishing ground in each part of the samples is kept basically consistent. The probability distribution and correlation of the various feature data is shown in fig. 3.

(5) 100 simple decision trees are adopted as sub-learning machines, and learning machine integration is carried out based on an AdaBoost algorithm.

And adopting a ten-fold cross validation method, selecting nine samples as training samples each time, and training the model. The remaining sample is used as a test sample to perform performance test and evaluation on the training result. Each time, a different sample was selected as a test sample, and training and evaluation were repeated 10 times, and the average of 10 evaluations was used as a model performance index.

The above ensemble learning model was used, the results averaged over ten runs: the accuracy is 0.82, and the confidence interval of the 95% confidence accuracy is 0.79-0.84. Results consistency test Kappa value was 0.64, achieving a better consistency level. For comparison, a strong learning machine of a C5.0 decision tree model is adopted, the accuracy of cross validation of ten folds is about 0.6-0.7, and the Kappa value of consistency test is 0.35. The integrated learning model with multiple environmental elements has better prediction accuracy and better generalization and popularization capability.

(6) The forecast is performed by taking the average data of the environment in 105-112 days in 2014 as an example. The forecast results are shown in fig. 4.

Claims (3)

1. An integrated learning fishery forecasting method utilizing ocean remote sensing multi-environment elements is characterized by comprising the following steps:

a1, dividing the obtained fishery environment and production data sample set into a plurality of samples according to a layered sampling method, wherein the number proportion of the samples with classification labels of a fishery and a non-fishery in each sample is kept consistent;

a2, adopting a decision tree with a simple structure as a sub-learning machine, wherein the boosting algorithm-based ensemble learning model comprises M mutually independent sub-learning machines, M is a positive integer greater than 1, the decision tree with the simple structure is a decision tree with one or two layers of nodes adopted by a decision stump, and the boosting ensemble method comprises the following three steps:

(1) initializing weight distribution Wm of training data, wherein for N learning samples, N is a positive integer greater than 1, and the weight of each training sample is initialized to W₁(i)＝1/N；

(2) Sub-learning machine G for training M bases in sequence_mAmong them are:

(i) learning by using Wm as weight distribution of training data set to obtain basic classifier G_m(X)，And calculates a classification error rate e thereof_mAnd classifier integration weight α_m：

$e_m=\underset{\{i:G_m\left(x_i\right)\&NotEqual;y_i\}}{\&Sigma;}{w}_m\left(i\right)$

${\mathrm{\&alpha;}}_m=\frac{1}{2}log\frac{1-e_m}{e_m}$

(ii) Updating weight distribution W of training data set_m+1：

$W_{m+1}\left(i\right)=\frac{W_m\left(i\right)}{Z_m}\exp (-{\mathrm{\&alpha;}}_m{y}_i{G}_m(x_i\left)\right)$

$Z_m=\underset{i=1}{\overset{N}{\&Sigma;}}{W}_m\left(i\right)\exp (-{\mathrm{\&alpha;}}_m{G}_{m\left(x_i\right)}{y}_i)$

Z_mIn order to normalize the factors, the method comprises the steps of,

(3) and weighted average is carried out on the results of all the sub-learning machines to be used as a final integrated learning model,

$G\left(x\right)=sign\left\{\underset{m=1}{\overset{M}{\&Sigma;}}{\mathrm{\&alpha;}}_m{G}_m\left(x\right)\right\}$

based on the model, fishery forecasting is carried out according to the input marine environment information.

2. The integrated learning fishery forecasting method using marine remote sensing multi-environment elements according to claim 1, wherein the step of obtaining a fishery environment and production data sample set comprises:

b1, taking the sea surface environment data obtained by the satellite remote sensing technology as model input data, and preprocessing the sea remote sensing multi-environment elements, wherein the preprocessing comprises the following steps:

cutting the acquired global sea surface environment data obtained by satellite remote sensing according to the latitude and longitude range of the region of interest to obtain a region of interest environment data subset;

interpolating numerical value loss at different time intervals and geographic positions due to the fact that remote sensing inversion data are influenced by factors such as weather conditions on the basis of neighborhood correlation;

b2, preprocessing fishery fishing production data, namely calculating unit fishing amount of effort according to fishery fishing record data, adopting a preset CPUE value as a threshold value for judging whether the fishing amount is a fishery or not according to the statistical distribution of the CPUE (unit fishing amount of effort) and fishery cost, establishing a classification label of the fishery, processing the record of the missing fishing amount information in the fishery data as an invalid sample,

wherein, CPUE is the yield of the fishery/the strength of the fishing effort;

b3, matching the fishery production data with the marine remote sensing environment, establishing a multi-feature set for constructing a model, and searching marine environment data closest to the space and time of the fishery production data according to the time and longitude and latitude of the fishery production data to serve as the matching environment features of the sample.

3. The integrated learning fishery forecasting method using marine remote sensing multi-environment elements according to claim 2, wherein the marine remote sensing multi-environment elements comprise chlorophyll a, Sea Surface Temperature (SST) and sea surface height.