1 of 25

Plantation Monitoring from Drone Images

1

Snehasis Mukherjee

Associate Professor

Shiv Nadar Institution of Eminence, Delhi NCR

2 of 25

Plantation Monitoring

Using Drones

​

Yashwanth Karumanchi (CVR College, Hyderabad)

Lakshmi Prasanna (WASSAN, Hyderabad)

Snehasis Mukherjee (...Shiv Nadar University…)

Nagesh Kolagani (...SSE, SIMATS Chennai…)

​

2

3 of 25

e-PRA: GIS-based Planning, Implementation, M&E

3

4 of 25

e-PRA: GIS-based Planning, Implementation, M&E

4

5 of 25

e-PRA: GIS-based Planning, Implementation, M&E

  1. Print GE satellite imagery of the village on a flexi sheet ~ 6’ x 4’
  2. Identify and mark land use (forest, community property resources, land terrain classification, cropping pattern, etc.) and private lands
  3. Community proposes various interventions using the flexi sheet
  4. Photograph and import these into QGIS (by a quasi-technical team)
  5. Visit in the field, gather details using GeoODK, import into QGIS
  6. Repeat / back to step 1: Print proposed GIS plan on a flexi-sheet, discuss, refine and finalize it in the Gram Sabha

5

6 of 25

Advantages of Drone Imagery over Satellite Imagery

6

  • At a time of our choice - Higher frequency - etc.
  • No cloud cover
  • With higher resolution
  • By local people
  • At less cost (?)

7 of 25

Drone Mapping - By Locals - Using Low-Cost Drones

Rural Technology Centres for Drone Mapping + Data Annotation:

  • Training programs: E.g. 25-27 June 2024 in Parigi near Hyderabad
  • For computer-literate rural youth
  • QGIS - GEarth - Drone Flying - WebOrthoDM - CVAnnotationT

7

8 of 25

Drone Mapping - By Locals - Using Low-Cost Drones

  • E.g. DJI mini 2 (and DJI mini 4 Pro):
    • RGB camera
    • Cost: USD $400 -$800 / Rs.(60k-120k)
    • Versus Rs. ~few lakhs for large quadcopters
      • Rs. 5.5 lakhs (?) for Garuda Aerospace’s 10L model
  • Flight height:
      • 15m-20m ✘ (e.g. crops)
        • Due to obstacles: Trees, Elec. lines, etc.
      • 50m-100m ✔ (e.g. plantations)
        • Accuracy: Reasonable ✔

8

9 of 25

Three Drone Sub-Projects In-Progress

  1. Latest base map preparation by locals using low-cost drones
    • As base map in our flexi sheet based exercises:

Planning - Implementation - Monitoring & Evaluation

      • Instead of using Google Earth imagery
        • Older
        • From less cropped summer months to avoid cloud cover
    • ℅ Sridhar, Lakshmi Prasanna, Pavan @ WASSAN
  • Plantation casualty monitoring
    • ℅ Yashwanth, Lakshmi Prasanna & Snehasis
  • Crop area estimation
    • ℅ Riya & Snehasis @ SNU

9

10 of 25

Drone Data

  • Near Hyderabad
    • In Vittalpur village, Parigi mandal,Vikarabad district
      • On 10-Aug-2023
      • ~50 ac.
      • Mainly plantations
      • Flight height: 50m
      • 253 images

10

11 of 25

Drone Data

11

12 of 25

Drone Data

12

13 of 25

Drone Data

13

14 of 25

Drone Data

14

15 of 25

Annotating Drone Imagery

15

16 of 25

Annotating Drone Imagery

    • 253 images x ~35-40 annotations / image
      • ~100 images: Lakshmi Prasanna
      • ~150 images: Outsourced: Cost: Rs. 1,000; Time: ~48 Hours
      • 9,534 annotations
        • 3 classes
          • Dead: 1,306
          • Stunted: 2,944
          • Good: 5,284
        • Randomly select same number images from each class (e.g. 1,206) and use for training and validation

16

17 of 25

Annotating Drone Imagery

  • Duplication error due to 80% overlap
    • Ortho Mosaicing will fix it ?
      • Reduces annotations: ~50 ac x ~60 trees/ac = ~3,000
      • Reduces image res. & hence classification accuracy ?

17

18 of 25

Detecting dead/stunted/good trees: (i) Classification

  • Using a DL/ML algorithm: CNN (‘Convolutional Neural Networks’)
    • Input: 1,206 images x 3 classes of trees
      • AlexNet: 71.6%
      • VGG19: 75.8%
      • GoogleNet: 77.1%
      • Efficient Net: 78.3%
      • VGG16: 78.9%

​

      • ResNet50: 95.6%
      • Xception: 97.1%

18

https://docs.google.com/document/d/19n3ec1Ofvm6Yd84gvBSxfs_rkk15iwErZDa9ChnfL44/edit?usp=sharing

19 of 25

Detecting dead/stunted/good trees: (ii) Object Detection

  • Using a DL/ML algorithm: YOLO (‘You Only Look Once’)
    • Free and Open Source Software (FOSS)
    • Version 8 (Jan 2023)
      • https://docs.ultralytics.com/
    • List of various ongoing experiments and their results
    • Current accuracy
      • 70% - 80%

19

20 of 25

Detecting dead/stunted/good trees: (ii) Object Detection

  • High resolution satellite imagery from Google Earth
    • 1-acre (Tirupati district)
    • 57.7%
  • Yolo’s pre-trained model

(℅ Rajesh; 23-July-2023)

    • 69.3%

20

21 of 25

Detecting dead/stunted/good trees: (ii) Object Detection

  • Drone imagery
    • 50 images: Split into 40 for training & 10 for validation randomly
    • Flexible (instead of fixed) boxes (℅ Snehasis; 6-Aug-2023)
    • v8s (instead of v8n)
    • 73%

​

21

22 of 25

Detecting dead/stunted/good trees: (ii) Object Detection

    • 255 images: 204 train and 51 val (random)
    • 77.6%

​

​

​

​

​

​

22

23 of 25

Detecting dead/stunted/good trees: (ii) Object Detection

    • Yolo-Version 9 (Feb 2023)
    • 255 images: 204 train and 51 val (random)
    • Flexible Boxes
    • v9 (Yolov9-c.pt)
    • Accuracy: 73%

​

​

​

​

​

​

23

24 of 25

ToDo: Detecting dead/stunted/good trees: (iii) Village Sy.

  • Map entire village / micro-watershed
    • ~1000-2000 acres
    • Multiple tree varieties
    • Large training data
      • Train and use rural youth

24

25 of 25

- Thank you -

25