DESC:TECHNICAL FIELD
[001] The present disclosure relates to species estimation, more particularly relates to a system and a method for estimating species density in a defined geographical location.

BACKGROUND
[002] Current species density estimations require conducting line transect surveys. However, the transect survey is highly manual and labour-intensive activity, and may span over a couple of months every year to determine species density.
[003] To overcome the drawbacks of the laborious task camera traps are being used to access remote areas within the forest reserves. These are cameras are activated by detecting movement in their scope of vision. The camera traps enable images of various animals/species moving within the reserve being captured. Further users and officials spend multiple hours trying to parse through these images and documents so that proper insights can be derived. However, there is a large gap between collection of camera trap images and them being processed to derive useful insights. This laborious process typically creates a lag of around 3-4 months from collection to insights.
[004] Hence there is a need in the art to decouple detection & classification of animal species from camera traps, and create a modular approach which is both robust and scalable.

SUMMARY
[005] In an implementation of the present disclosure a system for an automated pipeline for the classification and parsing of camera trap events by species class is disclosed. In accordance with the system a data fusion layer may integrate information from line-transect data, satellite images and patrolling data. Further a modified version of random encounter model is employed to calculate the point density estimates of the species using Bayesian estimates. The data analytics layer then create spatial-temporal dashboards, as well as on-demand reports for analysis. The system enables quick and timely flow of data and data-based insights from the beat officials to their supervisors with zero downtime.

BRIEF DESCRIPTION OF DRAWINGS
[006] The detailed description is set forth with reference to the accompanying drawings. The use of the same reference numerals may indicate similar or identical items. Various embodiments may utilize elements and/or components other than those illustrated in the drawings, and some elements and/or components may not be present in various embodiments. Elements and/or components in the figures are not necessarily drawn to scale. Throughout this disclosure, depending on the context, singular and plural terminology may be used interchangeably.
[007] Figure 1 illustrates an exemplary flow chart in accordance with the present disclosure.

DETAILED DESCRIPTION
[008] Some embodiments of the present disclosure, illustrating all its features, will now be discussed in detail. It must also be noted that as used herein and in the appended claims, the singular forms "a", "an" and "the" include plural references unless the context clearly dictates otherwise.
[009] In accordance with present disclosure a system and method for determining species density comprises creating a dataset and training an artificial intelligence model on the dataset created. Further to create a dataset, the present disclosure may rely on images captured using camera trap placed in the wild and an AI model based on iterative process using a dynamic human-in-the-loop system. The images captured may be run through a MegaDetector model. Further species detected using MegaDetector in the images are cropped out and placed in a single folder. Further the MegaDetector model is trained using PyTorch and the Yolov5 back-bone.
[0010] Further in accordance with the present disclosure the cropped images may be then transferred into an embedding space and an un-supervised algorithm to separate out the crops into plurality of different classes. Further a human-expert may assist to separate the images into different species classes.
[0011] Further the cropped images may be then trained on a SWIN Image Classifier. The SWIN transformer is a hierarchical transformer whose representations are calculated using a shifted window protocol. The use of a shifted window hierarchical transformer helps us to train a model with less data or even skewed datasets.
[0012] The separation of detection and classification approach as disclosed in the present exemplary embodiment work well, it also leads to a large number of false positives. This is largely because MDv5 and other models in this family. To overcome the issue of large number of false positives, the exemplary embodiment uses a single-shot model. The single-shot model enables detection and classification of animal species.
[0013] Further in accordance with the exemplary embodiment, Random Encounter Model may be used for camera trap to capture images of the species. The REM model in accordance with the embodiment may be further extended to the multi-species case in a Bayesian framework. Further REM model relies on the following formula for realizing the density:

[0014] Several camera trapping based methods exists viz. Random Encounter Model (REM), Random Encounter and Staying Time (REST) derived from REM employs video cameras, Distance Sampling with Camera Traps (CT-DS). Random Encounter Model is widely used density estimate model for camera trap data as it has no need for species- specific study design. REM is based on modelling random encounters between moving animals and static camera traps, taking into account key variables such as camera detection zone, defined by its radius and angle, and the daily distance that affect the encounter rate. The main advantage of REM is that individual identification is not needed, so then REM can be used to monitor both unmarked and marked populations without the need of capture and tag animals. More than one species can be potentially monitored during the same survey.
[0015] Random Encounter and Staying Time (REST) is extension of the REM. The REST model describes the relationship among staying time, trapping rate, and density, which is estimable using a frequentist or Bayesian approach. REST requires camera trap takes a series of pictures at very short intervals (e.g. 0.1 seconds) for entire survey period or the video of an animal is taken when it is in front of the camera. Distance Sampling with Camera Traps (CT-DS) is used when camera trap images are recorded at a constant rate as long as animal is present in front of the camera. Based on horizontal radial distance between camera and animal (h) and angle and estimated probability of obtaining an image of an animal that is within h and w where w is the truncation distance, the populatiovarn density is calculate d. Due to these reasons REM is suitable for current scenario to get density estimates of animals.
[0016] In accordance with exemplary embodiment the method may comprise of a dataset creation. The dataset creation may further comprise curating images captured from the camera trap positioned at various places throughout the defined geographical location. Further the captured images may be processed using a MegaDetector v5 model. The MegaDetector v5 model may be configured to detect and generate distinct clusters for animals, humans and vehicles. Further clusters annotator is included for identifying the species of each of the animal crops present in the dataset.
[0017] The method further comprises training a model. The training dataset may be prepared and used to train a SWIN ImageClassifier. The SWIN ImageClassifier is configured to provide a first layer model. This may be used to run inference on a larger dataset consisting of more than camera trap imagery. Further all the images may be annotated for cleaning the different classes to create a larger dataset of species crops, which can then be used to train smaller quantization-aware models.
[0018] The method further comprises creating quantized models, by exporting the weights to Torchscript or Int8 models, or by using quantization-aware (QA) training. Further for training a TFJS model, the dataset modified such that training can be used to train a TensorFlow model. This is then used to train lighter models. And create a model zoo. The model zoo may provide similar or better accuracy than the chained approach of a MDv5 and a SWIN ImageClassifier. Further the model zoo may be trained using a different model architecture on a large crop of species.
[0019] The method as disclosed further comprises a reporting and data fusion layer. The reporting and data fusion layer is configured to receive data from the model zoo. The model zoo is configured to identify species in each of the camera trap images. The model zoo may be then implemented in a modified version of the Random Encounter Model.
[0020] The random encounters between moving animals and static camera traps provide an encounter rate. The key variables that affect the encounter rate are camera detection zone (defined by the radius and angle of detection). The present method enables no species-specific or individual identification is required for calculating the densities. The data is piped into the data fusion layer, which creates spatio-temporal maps of the wildlife sanctuary. These maps and the reports that are generated are useful for understanding the species distribution in the sanctuary. Human-Animal conflict zones and potential eco-tourism hotspots can be generated using the data collected.
[0021] Figure 1 illustrates an exemplary flow chart in accordance with the present disclosure. The method as disclosed comprises a step 102, for creating a data set. Further at step 104, the method comprises training an AI model, and at step 106, comprises density estimation of a species.
[0022] The step 102, further comprises at step 108 capturing using a REM camera trap, images of various species detected in by the camera trap. Further at step 110, processing the images using a MegaDetector v5 model. Further at step 112, generating a plurality set of cropped images. The plurality sets of images may comprise images of animals, humans, vehicles and others. At step 114, encoding the plurality sets of images into a vector representation of the image data.
[0023] Further at step 116, identifying various clusters of animals, humans, vehicles and others using an unsupervised clustering algorithm. The clustering algorithm may have a k value of 30. The unsupervised clustering may be configured to cluster into thirty distinct classes of data.
[0024] In accordance with the exemplary embodiment at step 118, the method comprises annotating and identifying species of each of the animal present in the dataset. Further at step 120, augmenting the dataset with image manipulations.
[0025] Further at step 122, training a SWIN ImageClassifier using the dataset to obtain a first layer model. At step 124, running the first layer model on a larger dataset, and annotating the lager dataset to train smaller quantization-aware models. Quantization refers to the techniques for using 8-bit (lower bandwidth) integer instructions to reduce the model size and run faster instructions. The use of lower precision data enables performance gains by reducing (a) 4x reduction in model size;(b) 2x-4x reduction in memory bandwidth. Both improve inference speed by 2x-4x times. For reducing model bandwidth, we have used TensorflowJS, ONNX and tiny YOLOv5, a deep learning toolkit built using PyTorch. There are two methods for creating quantized models - (1) By exporting the weights to Torchscript or Int8 models or (2) By using quantization-aware(QA) training. While the former uses post-quantization techniques, QA training emulates inference-time quantization during the training so that the resulting lower-precision model can benefit during deployment. For training a TFJS model, one needs to modify the dataset such that it can be used to train a TensorFlow model. This is then used to train lighter models. And create what we call as a model zoo. Which provides similar or better accuracy than the chained approach of a MDv5 and a SWIN ImageClassifier.
[0026] In this second stage, a model zoo is obtained, which is trained using different model architectures on a large crop of species. This makes our system robust to biases or errors induced due to the choice of model architecture. The size of the models is also less than 40MB, and they are single-shot detectors and classifiers. This approach is computationally efficient and hence is also energy efficient. These models can therefore be deployed either in combination or as separate model files on devices.
[0027] Furter in accordance with the exemplary embodiment, the step 106 may further comprise a step 126, identifying using a model zoo, animal species in each of the camera trap images. Further at step 128, applying Random Encounter Model, having Bayesian approach to for calculating the densities. The Bayesian approach helps with abundance estimation by assuming that the distribution of camera traps is random. It models the random encounters between moving animals and static camera traps to provide an encounter rate. The key variables that affect the encounter rate are camera detection zone (defined by the radius and angle of detection). The main advantage is that no species-specific or individual identification is required.
[0028] Further at step 130, fusing data into a data fusion layer, to creates spatio-temporal maps of the wildlife sanctuary.
,CLAIMS:1. A method for estimating species density in a defined geographical location, the method comprises:
creating a data set, wherein creating dataset further comprises:
capturing using a REM camera trap, images of various objects detected by the camera trap;
processing the images using a MegaDetector v5 model;
training an AI model, wherein training further comprises;
training a SWIN ImageClassifier using the dataset to obtain a first layer model;
estimating species density, wherein estimating further comprises:
identifying using a model zoo, animal species in each of the camera trap images;
applying Random Encounter Model, having Bayesian approach to for calculating the densities; and
fusing data into a data fusion layer, to creates spatio-temporal maps of the wildlife sanctuary.

2. The method as claimed in claim 1, wherein creating dataset further comprises, generating a plurality set of cropped images.

3. The method as claimed in claim 1, wherein creating dataset further comprises, encoding the plurality sets of images into a vector representation of the image data.

4. The method as claimed in claim 1, wherein creating dataset further comprises, identifying various clusters of animals, humans, vehicles and others using an unsupervised clustering algorithm.

5. The method as claimed in claim 1, wherein creating dataset further comprises, annotating and identifying species of each of the animal present in the dataset.

6. The method as claimed in claim 1, wherein creating dataset further comprises, augmenting the dataset with image manipulations.

7. The method as claimed in claim 1, wherein training further comprises, running the first layer model on a larger dataset, and annotating the lager dataset to train smaller quantization-aware models.