This project focuses on using satellite imagery to understand environmental conditions in great detail. We're using advanced computer vision techniques to analyze this imagery, looking for signs of drought stress. By training our algorithms on a carefully labeled dataset, we aim to identify patterns that indicate if droughts are likely to occur.

Dataset

The current dataset consists of 86,317 train and 10,778 validation satellite images, 65x65 pixels each, in 10 spectrum bands, with 10,774 images withheld to test long-term generalization (107,869 total). Human experts (pastoralists) have labeled these with the number of cows that the geographic location at the center of the image could support (0, 1, 2, or 3+ cows). Each pixel represents a 30 meter square, so the images at full size are 1.95 kilometers across. Pastoralists are asked to rate the quality of the area within 20 meters of where they are standing, which corresponds to an area slightly larger a single pixel. Since forage quality is correlated across space, the larger image may be useful for prediction.

The data is in TFRecords format, split into train and val, and takes up ~4.3GB (2.15GB zipped). You can learn more about the format of the satellite images here.

The data used in this research was collected through a research collaboration between the International Livestock Research Institute, Cornell University, and UC San Diego. It was supported by the Atkinson Centre for a Sustainable Future’s Academic Venture Fund, Australian Aid through the AusAID Development Research Awards Scheme Agreement No. 66138, the National Science Foundation (0832782, 1059284, 1522054), and ARO grant W911-NF-14-1-0498.

Downloading the Data

In [ ]:
!echo "Downloading data"
!curl -SL https://storage.googleapis.com/wandb_datasets/dw_train_86K_val_10K.zip > dw_data.zip
!unzip dw_data.zip
!rm dw_data.zip
!mv droughtwatch_data/ data/

Downloading data
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 2050M  100 2050M    0     0  86.6M      0  0:00:23  0:00:23 --:--:--  118M
Archive:  dw_data.zip
   creating: droughtwatch_data/
   creating: droughtwatch_data/val/
  inflating: droughtwatch_data/val/part-r-00090  
  inflating: droughtwatch_data/val/part-r-00061  
  inflating: droughtwatch_data/val/part-r-00052  
  inflating: droughtwatch_data/val/part-r-00043  
  inflating: droughtwatch_data/val/part-r-00040  
  inflating: droughtwatch_data/val/part-r-00042  
  inflating: droughtwatch_data/val/part-r-00067  
  inflating: droughtwatch_data/val/part-r-00026  
  inflating: droughtwatch_data/val/part-r-00046  
  inflating: droughtwatch_data/val/part-r-00023  
  inflating: droughtwatch_data/val/part-r-00083  
  inflating: droughtwatch_data/val/part-r-00011  
  inflating: droughtwatch_data/val/part-r-00058  
  inflating: droughtwatch_data/val/part-r-00012  
  inflating: droughtwatch_data/val/part-r-00078  
  inflating: droughtwatch_data/val/part-r-00082  
  inflating: droughtwatch_data/val/part-r-00038  
  inflating: droughtwatch_data/val/part-r-00071  
  inflating: droughtwatch_data/val/part-r-00088  
  inflating: droughtwatch_data/val/part-r-00029  
  inflating: droughtwatch_data/val/part-r-00097  
  inflating: droughtwatch_data/val/part-r-00096  
  inflating: droughtwatch_data/val/part-r-00077  
  inflating: droughtwatch_data/val/part-r-00016  
  inflating: droughtwatch_data/val/part-r-00076  
  inflating: droughtwatch_data/val/part-r-00057  
  inflating: droughtwatch_data/val/part-r-00009  
  inflating: droughtwatch_data/val/part-r-00054  
  inflating: droughtwatch_data/val/part-r-00028  
  inflating: droughtwatch_data/val/part-r-00014  
  inflating: droughtwatch_data/val/part-r-00001  
  inflating: droughtwatch_data/val/part-r-00080  
  inflating: droughtwatch_data/val/part-r-00013  
  inflating: droughtwatch_data/val/part-r-00004  
  inflating: droughtwatch_data/val/part-r-00070  
  inflating: droughtwatch_data/val/part-r-00037  
  inflating: droughtwatch_data/val/part-r-00005  
  inflating: droughtwatch_data/val/part-r-00010  
  inflating: droughtwatch_data/val/part-r-00024  
  inflating: droughtwatch_data/val/part-r-00075  
  inflating: droughtwatch_data/val/part-r-00074  
  inflating: droughtwatch_data/val/part-r-00017  
  inflating: droughtwatch_data/val/part-r-00055  
  inflating: droughtwatch_data/val/part-r-00091  
  inflating: droughtwatch_data/val/part-r-00020  
  inflating: droughtwatch_data/val/part-r-00095  
  inflating: droughtwatch_data/val/part-r-00079  
  inflating: droughtwatch_data/val/part-r-00047  
  inflating: droughtwatch_data/val/part-r-00051  
  inflating: droughtwatch_data/val/part-r-00073  
  inflating: droughtwatch_data/val/part-r-00053  
  inflating: droughtwatch_data/val/part-r-00060  
  inflating: droughtwatch_data/val/part-r-00086  
  inflating: droughtwatch_data/val/part-r-00048  
  inflating: droughtwatch_data/val/part-r-00006  
  inflating: droughtwatch_data/val/part-r-00059  
  inflating: droughtwatch_data/val/part-r-00002  
  inflating: droughtwatch_data/val/part-r-00094  
  inflating: droughtwatch_data/val/part-r-00033  
  inflating: droughtwatch_data/val/part-r-00066  
  inflating: droughtwatch_data/val/part-r-00015  
  inflating: droughtwatch_data/val/part-r-00000  
  inflating: droughtwatch_data/val/part-r-00050  
  inflating: droughtwatch_data/val/part-r-00093  
  inflating: droughtwatch_data/val/part-r-00035  
  inflating: droughtwatch_data/val/part-r-00098  
  inflating: droughtwatch_data/val/part-r-00099  
  inflating: droughtwatch_data/val/part-r-00084  
  inflating: droughtwatch_data/val/part-r-00003  
  inflating: droughtwatch_data/val/part-r-00032  
  inflating: droughtwatch_data/val/part-r-00018  
  inflating: droughtwatch_data/val/part-r-00062  
  inflating: droughtwatch_data/val/part-r-00045  
  inflating: droughtwatch_data/val/part-r-00089  
  inflating: droughtwatch_data/val/part-r-00087  
  inflating: droughtwatch_data/val/part-r-00027  
  inflating: droughtwatch_data/val/part-r-00064  
  inflating: droughtwatch_data/val/part-r-00085  
  inflating: droughtwatch_data/val/part-r-00025  
 extracting: droughtwatch_data/val/_SUCCESS  
  inflating: droughtwatch_data/val/part-r-00030  
  inflating: droughtwatch_data/val/part-r-00065  
  inflating: droughtwatch_data/val/part-r-00034  
  inflating: droughtwatch_data/val/part-r-00031  
  inflating: droughtwatch_data/val/part-r-00007  
  inflating: droughtwatch_data/val/part-r-00008  
  inflating: droughtwatch_data/val/part-r-00041  
  inflating: droughtwatch_data/val/part-r-00063  
  inflating: droughtwatch_data/val/part-r-00069  
  inflating: droughtwatch_data/val/part-r-00021  
  inflating: droughtwatch_data/val/part-r-00044  
  inflating: droughtwatch_data/val/part-r-00081  
  inflating: droughtwatch_data/val/part-r-00039  
  inflating: droughtwatch_data/val/part-r-00068  
  inflating: droughtwatch_data/val/part-r-00056  
  inflating: droughtwatch_data/val/part-r-00072  
  inflating: droughtwatch_data/val/part-r-00019  
  inflating: droughtwatch_data/val/part-r-00036  
  inflating: droughtwatch_data/val/part-r-00092  
  inflating: droughtwatch_data/val/part-r-00049  
  inflating: droughtwatch_data/val/part-r-00022  
   creating: droughtwatch_data/train/
  inflating: droughtwatch_data/train/part-r-01174  
  inflating: droughtwatch_data/train/part-r-01012  
  inflating: droughtwatch_data/train/part-r-00110  
  inflating: droughtwatch_data/train/part-r-00132  
  inflating: droughtwatch_data/train/part-r-01168  
  inflating: droughtwatch_data/train/part-r-01011  
  inflating: droughtwatch_data/train/part-r-01084  
  inflating: droughtwatch_data/train/part-r-01138  
  inflating: droughtwatch_data/train/part-r-00090  
  inflating: droughtwatch_data/train/part-r-00167  
  inflating: droughtwatch_data/train/part-r-00128  
  inflating: droughtwatch_data/train/part-r-01060  
  inflating: droughtwatch_data/train/part-r-00061  
  inflating: droughtwatch_data/train/part-r-00116  
  inflating: droughtwatch_data/train/part-r-00052  
  inflating: droughtwatch_data/train/part-r-01091  
  inflating: droughtwatch_data/train/part-r-01007  
  inflating: droughtwatch_data/train/part-r-00043  
  inflating: droughtwatch_data/train/part-r-01120  
  inflating: droughtwatch_data/train/part-r-01025  
  inflating: droughtwatch_data/train/part-r-01056  
  inflating: droughtwatch_data/train/part-r-00040  
  inflating: droughtwatch_data/train/part-r-00178  
  inflating: droughtwatch_data/train/part-r-01136  
  inflating: droughtwatch_data/train/part-r-00106  
  inflating: droughtwatch_data/train/part-r-01066  
  inflating: droughtwatch_data/train/part-r-01104  
  inflating: droughtwatch_data/train/part-r-00042  
  inflating: droughtwatch_data/train/part-r-01192  
  inflating: droughtwatch_data/train/part-r-00136  
  inflating: droughtwatch_data/train/part-r-00067  
  inflating: droughtwatch_data/train/part-r-01181  
  inflating: droughtwatch_data/train/part-r-00190  
  inflating: droughtwatch_data/train/part-r-00196  
  inflating: droughtwatch_data/train/part-r-00026  
  inflating: droughtwatch_data/train/part-r-00177  
  inflating: droughtwatch_data/train/part-r-01039  
  inflating: droughtwatch_data/train/part-r-00046  
  inflating: droughtwatch_data/train/part-r-01024  
  inflating: droughtwatch_data/train/part-r-01102  
  inflating: droughtwatch_data/train/part-r-00023  
  inflating: droughtwatch_data/train/part-r-01188  
  inflating: droughtwatch_data/train/part-r-00184  
  inflating: droughtwatch_data/train/part-r-00137  
  inflating: droughtwatch_data/train/part-r-01199  
  inflating: droughtwatch_data/train/part-r-01115  
  inflating: droughtwatch_data/train/part-r-00083  
  inflating: droughtwatch_data/train/part-r-01014  
  inflating: droughtwatch_data/train/part-r-00011  
  inflating: droughtwatch_data/train/part-r-00144  
  inflating: droughtwatch_data/train/part-r-00058  
  inflating: droughtwatch_data/train/part-r-00115  
  inflating: droughtwatch_data/train/part-r-01154  
  inflating: droughtwatch_data/train/part-r-00012  
  inflating: droughtwatch_data/train/part-r-00156  
  inflating: droughtwatch_data/train/part-r-00078  
  inflating: droughtwatch_data/train/part-r-01026  
  inflating: droughtwatch_data/train/part-r-00082  
  inflating: droughtwatch_data/train/part-r-01171  
  inflating: droughtwatch_data/train/part-r-00153  
  inflating: droughtwatch_data/train/part-r-00172  
  inflating: droughtwatch_data/train/part-r-01037  
  inflating: droughtwatch_data/train/part-r-00165  
  inflating: droughtwatch_data/train/part-r-01022  
  inflating: droughtwatch_data/train/part-r-00150  
  inflating: droughtwatch_data/train/part-r-01095  
  inflating: droughtwatch_data/train/part-r-00108  
  inflating: droughtwatch_data/train/part-r-00038  
  inflating: droughtwatch_data/train/part-r-00119  
  inflating: droughtwatch_data/train/part-r-01170  
  inflating: droughtwatch_data/train/part-r-01079  
  inflating: droughtwatch_data/train/part-r-00071  
  inflating: droughtwatch_data/train/part-r-01076  
  inflating: droughtwatch_data/train/part-r-00088  
  inflating: droughtwatch_data/train/part-r-01090  
  inflating: droughtwatch_data/train/part-r-00112  
  inflating: droughtwatch_data/train/part-r-00029  
  inflating: droughtwatch_data/train/part-r-00097  
  inflating: droughtwatch_data/train/part-r-01092  
  inflating: droughtwatch_data/train/part-r-01054  
  inflating: droughtwatch_data/train/part-r-01001  
  inflating: droughtwatch_data/train/part-r-01198  
  inflating: droughtwatch_data/train/part-r-00174  
  inflating: droughtwatch_data/train/part-r-01109  
  inflating: droughtwatch_data/train/part-r-01161  
  inflating: droughtwatch_data/train/part-r-01125  
  inflating: droughtwatch_data/train/part-r-01128  
  inflating: droughtwatch_data/train/part-r-01019  
  inflating: droughtwatch_data/train/part-r-00096  
  inflating: droughtwatch_data/train/part-r-01195  
  inflating: droughtwatch_data/train/part-r-00077  
  inflating: droughtwatch_data/train/part-r-00016  
  inflating: droughtwatch_data/train/part-r-00121  
  inflating: droughtwatch_data/train/part-r-01087  
  inflating: droughtwatch_data/train/part-r-00199  
  inflating: droughtwatch_data/train/part-r-00185  
  inflating: droughtwatch_data/train/part-r-01116  
  inflating: droughtwatch_data/train/part-r-01101  
  inflating: droughtwatch_data/train/part-r-01178  
  inflating: droughtwatch_data/train/part-r-01035  
  inflating: droughtwatch_data/train/part-r-01006  
  inflating: droughtwatch_data/train/part-r-01183  
  inflating: droughtwatch_data/train/part-r-00104  
  inflating: droughtwatch_data/train/part-r-01155  
  inflating: droughtwatch_data/train/part-r-01061  
  inflating: droughtwatch_data/train/part-r-01008  
  inflating: droughtwatch_data/train/part-r-01063  
  inflating: droughtwatch_data/train/part-r-01137  
  inflating: droughtwatch_data/train/part-r-01034  
  inflating: droughtwatch_data/train/part-r-00171  
  inflating: droughtwatch_data/train/part-r-01159  
  inflating: droughtwatch_data/train/part-r-00109  
  inflating: droughtwatch_data/train/part-r-00076  
  inflating: droughtwatch_data/train/part-r-01082  
  inflating: droughtwatch_data/train/part-r-01004  
  inflating: droughtwatch_data/train/part-r-01145  
  inflating: droughtwatch_data/train/part-r-00161  
  inflating: droughtwatch_data/train/part-r-00101  
  inflating: droughtwatch_data/train/part-r-00057  
  inflating: droughtwatch_data/train/part-r-01182  
  inflating: droughtwatch_data/train/part-r-01118  
  inflating: droughtwatch_data/train/part-r-00168  
  inflating: droughtwatch_data/train/part-r-00102  
  inflating: droughtwatch_data/train/part-r-01085  
  inflating: droughtwatch_data/train/part-r-00183  
  inflating: droughtwatch_data/train/part-r-01187  
  inflating: droughtwatch_data/train/part-r-00009  
  inflating: droughtwatch_data/train/part-r-01114  
  inflating: droughtwatch_data/train/part-r-01051  
  inflating: droughtwatch_data/train/part-r-01185  
  inflating: droughtwatch_data/train/part-r-01057  
  inflating: droughtwatch_data/train/part-r-01119  
  inflating: droughtwatch_data/train/part-r-00189  
  inflating: droughtwatch_data/train/part-r-01081  
  inflating: droughtwatch_data/train/part-r-00054  
  inflating: droughtwatch_data/train/part-r-01144  
  inflating: droughtwatch_data/train/part-r-00179  
  inflating: droughtwatch_data/train/part-r-01196  
  inflating: droughtwatch_data/train/part-r-01072  
  inflating: droughtwatch_data/train/part-r-01172  
  inflating: droughtwatch_data/train/part-r-01156  
  inflating: droughtwatch_data/train/part-r-01176  
  inflating: droughtwatch_data/train/part-r-00028  
  inflating: droughtwatch_data/train/part-r-00014  
  inflating: droughtwatch_data/train/part-r-00001  
  inflating: droughtwatch_data/train/part-r-01062  
  inflating: droughtwatch_data/train/part-r-01107  
  inflating: droughtwatch_data/train/part-r-00145  
  inflating: droughtwatch_data/train/part-r-00080  
  inflating: droughtwatch_data/train/part-r-00124  
  inflating: droughtwatch_data/train/part-r-01151  
  inflating: droughtwatch_data/train/part-r-00176  
  inflating: droughtwatch_data/train/part-r-00013  
  inflating: droughtwatch_data/train/part-r-00004  
  inflating: droughtwatch_data/train/part-r-00070  
  inflating: droughtwatch_data/train/part-r-00157  
  inflating: droughtwatch_data/train/part-r-01113  
  inflating: droughtwatch_data/train/part-r-00037  
  inflating: droughtwatch_data/train/part-r-01173  
  inflating: droughtwatch_data/train/part-r-00005  
  inflating: droughtwatch_data/train/part-r-01158  
  inflating: droughtwatch_data/train/part-r-01164  
  inflating: droughtwatch_data/train/part-r-01175  
  inflating: droughtwatch_data/train/part-r-00148  
  inflating: droughtwatch_data/train/part-r-01049  
  inflating: droughtwatch_data/train/part-r-01106  
  inflating: droughtwatch_data/train/part-r-01002  
  inflating: droughtwatch_data/train/part-r-00142  
  inflating: droughtwatch_data/train/part-r-00126  
  inflating: droughtwatch_data/train/part-r-00010  
  inflating: droughtwatch_data/train/part-r-00024  
  inflating: droughtwatch_data/train/part-r-01184  
  inflating: droughtwatch_data/train/part-r-01193  
  inflating: droughtwatch_data/train/part-r-00075  
  inflating: droughtwatch_data/train/part-r-00175  
  inflating: droughtwatch_data/train/part-r-01160  
  inflating: droughtwatch_data/train/part-r-01163  
  inflating: droughtwatch_data/train/part-r-00074  
  inflating: droughtwatch_data/train/part-r-01169  
  inflating: droughtwatch_data/train/part-r-01110  
  inflating: droughtwatch_data/train/part-r-00017  
  inflating: droughtwatch_data/train/part-r-01099  
  inflating: droughtwatch_data/train/part-r-00055  
  inflating: droughtwatch_data/train/part-r-00166  
  inflating: droughtwatch_data/train/part-r-00091  
  inflating: droughtwatch_data/train/part-r-00020  
  inflating: droughtwatch_data/train/part-r-01010  
  inflating: droughtwatch_data/train/part-r-01179  
  inflating: droughtwatch_data/train/part-r-01052  
  inflating: droughtwatch_data/train/part-r-00133  
  inflating: droughtwatch_data/train/part-r-00120  
  inflating: droughtwatch_data/train/part-r-00095  
  inflating: droughtwatch_data/train/part-r-01059  
  inflating: droughtwatch_data/train/part-r-00134  
  inflating: droughtwatch_data/train/part-r-01003  
  inflating: droughtwatch_data/train/part-r-01157  
  inflating: droughtwatch_data/train/part-r-00173  
  inflating: droughtwatch_data/train/part-r-01141  
  inflating: droughtwatch_data/train/part-r-00079  
  inflating: droughtwatch_data/train/part-r-00047  
  inflating: droughtwatch_data/train/part-r-00051  
  inflating: droughtwatch_data/train/part-r-00073  
  inflating: droughtwatch_data/train/part-r-01080  
  inflating: droughtwatch_data/train/part-r-01015  
  inflating: droughtwatch_data/train/part-r-00122  
  inflating: droughtwatch_data/train/part-r-01027  
  inflating: droughtwatch_data/train/part-r-00162  
  inflating: droughtwatch_data/train/part-r-00053  
  inflating: droughtwatch_data/train/part-r-00111  
  inflating: droughtwatch_data/train/part-r-00152  
  inflating: droughtwatch_data/train/part-r-00060  
  inflating: droughtwatch_data/train/part-r-00086  
  inflating: droughtwatch_data/train/part-r-01041  
  inflating: droughtwatch_data/train/part-r-00048  
  inflating: droughtwatch_data/train/part-r-01055  
  inflating: droughtwatch_data/train/part-r-00006  
  inflating: droughtwatch_data/train/part-r-00059  
  inflating: droughtwatch_data/train/part-r-01140  
  inflating: droughtwatch_data/train/part-r-01165  
  inflating: droughtwatch_data/train/part-r-01065  
  inflating: droughtwatch_data/train/part-r-01086  
  inflating: droughtwatch_data/train/part-r-01048  
  inflating: droughtwatch_data/train/part-r-01045  
  inflating: droughtwatch_data/train/part-r-01126  
  inflating: droughtwatch_data/train/part-r-00002  
  inflating: droughtwatch_data/train/part-r-01009  
  inflating: droughtwatch_data/train/part-r-01103  
  inflating: droughtwatch_data/train/part-r-00131  
  inflating: droughtwatch_data/train/part-r-00094  
  inflating: droughtwatch_data/train/part-r-01098  
  inflating: droughtwatch_data/train/part-r-00033  
  inflating: droughtwatch_data/train/part-r-01177  
  inflating: droughtwatch_data/train/part-r-01148  
  inflating: droughtwatch_data/train/part-r-00123  
  inflating: droughtwatch_data/train/part-r-00100  
  inflating: droughtwatch_data/train/part-r-01197  
  inflating: droughtwatch_data/train/part-r-00113  
  inflating: droughtwatch_data/train/part-r-00066  
  inflating: droughtwatch_data/train/part-r-01053  
  inflating: droughtwatch_data/train/part-r-01071  
  inflating: droughtwatch_data/train/part-r-00015  
  inflating: droughtwatch_data/train/part-r-00000  
  inflating: droughtwatch_data/train/part-r-01143  
  inflating: droughtwatch_data/train/part-r-00140  
  inflating: droughtwatch_data/train/part-r-00050  
  inflating: droughtwatch_data/train/part-r-00093  
  inflating: droughtwatch_data/train/part-r-01040  
  inflating: droughtwatch_data/train/part-r-01097  
  inflating: droughtwatch_data/train/part-r-01112  
  inflating: droughtwatch_data/train/part-r-01149  
  inflating: droughtwatch_data/train/part-r-01023  
  inflating: droughtwatch_data/train/part-r-00035  
  inflating: droughtwatch_data/train/part-r-01121  
  inflating: droughtwatch_data/train/part-r-00118  
  inflating: droughtwatch_data/train/part-r-00158  
  inflating: droughtwatch_data/train/part-r-00146  
  inflating: droughtwatch_data/train/part-r-00098  
  inflating: droughtwatch_data/train/part-r-00192  
  inflating: droughtwatch_data/train/part-r-01152  
  inflating: droughtwatch_data/train/part-r-00125  
  inflating: droughtwatch_data/train/part-r-00099  
  inflating: droughtwatch_data/train/part-r-00084  
  inflating: droughtwatch_data/train/part-r-01028  
  inflating: droughtwatch_data/train/part-r-01046  
  inflating: droughtwatch_data/train/part-r-01180  
  inflating: droughtwatch_data/train/part-r-00003  
  inflating: droughtwatch_data/train/part-r-01100  
  inflating: droughtwatch_data/train/part-r-00180  
  inflating: droughtwatch_data/train/part-r-00139  
  inflating: droughtwatch_data/train/part-r-00160  
  inflating: droughtwatch_data/train/part-r-01005  
  inflating: droughtwatch_data/train/part-r-01075  
  inflating: droughtwatch_data/train/part-r-00032  
  inflating: droughtwatch_data/train/part-r-01067  
  inflating: droughtwatch_data/train/part-r-00151  
  inflating: droughtwatch_data/train/part-r-00018  
  inflating: droughtwatch_data/train/part-r-00143  
  inflating: droughtwatch_data/train/part-r-01133  
  inflating: droughtwatch_data/train/part-r-01135  
  inflating: droughtwatch_data/train/part-r-01017  
  inflating: droughtwatch_data/train/part-r-00187  
  inflating: droughtwatch_data/train/part-r-00129  
  inflating: droughtwatch_data/train/part-r-01020  
  inflating: droughtwatch_data/train/part-r-00062  
  inflating: droughtwatch_data/train/part-r-00045  
  inflating: droughtwatch_data/train/part-r-00147  
  inflating: droughtwatch_data/train/part-r-01064  
  inflating: droughtwatch_data/train/part-r-00089  
  inflating: droughtwatch_data/train/part-r-01074  
  inflating: droughtwatch_data/train/part-r-00135  
  inflating: droughtwatch_data/train/part-r-01142  
  inflating: droughtwatch_data/train/part-r-01036  
  inflating: droughtwatch_data/train/part-r-01078  
  inflating: droughtwatch_data/train/part-r-01050  
  inflating: droughtwatch_data/train/part-r-01166  
  inflating: droughtwatch_data/train/part-r-01134  
  inflating: droughtwatch_data/train/part-r-00191  
  inflating: droughtwatch_data/train/part-r-00182  
  inflating: droughtwatch_data/train/part-r-01073  
  inflating: droughtwatch_data/train/part-r-00087  
  inflating: droughtwatch_data/train/part-r-01029  
  inflating: droughtwatch_data/train/part-r-00027  
  inflating: droughtwatch_data/train/part-r-01018  
  inflating: droughtwatch_data/train/part-r-00198  
  inflating: droughtwatch_data/train/part-r-00064  
  inflating: droughtwatch_data/train/part-r-00149  
  inflating: droughtwatch_data/train/part-r-00193  
  inflating: droughtwatch_data/train/part-r-00154  
  inflating: droughtwatch_data/train/part-r-01162  
  inflating: droughtwatch_data/train/part-r-00085  
  inflating: droughtwatch_data/train/part-r-00025  
  inflating: droughtwatch_data/train/part-r-01117  
  inflating: droughtwatch_data/train/part-r-00130  
  inflating: droughtwatch_data/train/part-r-01124  
 extracting: droughtwatch_data/train/_SUCCESS  
  inflating: droughtwatch_data/train/part-r-01189  
  inflating: droughtwatch_data/train/part-r-01032  
  inflating: droughtwatch_data/train/part-r-00030  
  inflating: droughtwatch_data/train/part-r-01068  
  inflating: droughtwatch_data/train/part-r-00065  
  inflating: droughtwatch_data/train/part-r-01108  
  inflating: droughtwatch_data/train/part-r-00141  
  inflating: droughtwatch_data/train/part-r-00138  
  inflating: droughtwatch_data/train/part-r-01096  
  inflating: droughtwatch_data/train/part-r-01089  
  inflating: droughtwatch_data/train/part-r-01167  
  inflating: droughtwatch_data/train/part-r-01150  
  inflating: droughtwatch_data/train/part-r-01038  
  inflating: droughtwatch_data/train/part-r-01030  
  inflating: droughtwatch_data/train/part-r-00034  
  inflating: droughtwatch_data/train/part-r-01083  
  inflating: droughtwatch_data/train/part-r-00031  
  inflating: droughtwatch_data/train/part-r-00164  
  inflating: droughtwatch_data/train/part-r-00197  
  inflating: droughtwatch_data/train/part-r-00007  
  inflating: droughtwatch_data/train/part-r-00103  
  inflating: droughtwatch_data/train/part-r-01094  
  inflating: droughtwatch_data/train/part-r-01058  
  inflating: droughtwatch_data/train/part-r-01069  
  inflating: droughtwatch_data/train/part-r-00105  
  inflating: droughtwatch_data/train/part-r-01000  
  inflating: droughtwatch_data/train/part-r-01093  
  inflating: droughtwatch_data/train/part-r-01139  
  inflating: droughtwatch_data/train/part-r-00008  
  inflating: droughtwatch_data/train/part-r-00195  
  inflating: droughtwatch_data/train/part-r-01033  
  inflating: droughtwatch_data/train/part-r-01127  
  inflating: droughtwatch_data/train/part-r-01147  
  inflating: droughtwatch_data/train/part-r-00186  
  inflating: droughtwatch_data/train/part-r-00155  
  inflating: droughtwatch_data/train/part-r-00041  
  inflating: droughtwatch_data/train/part-r-01043  
  inflating: droughtwatch_data/train/part-r-00194  
  inflating: droughtwatch_data/train/part-r-00114  
  inflating: droughtwatch_data/train/part-r-01132  
  inflating: droughtwatch_data/train/part-r-01153  
  inflating: droughtwatch_data/train/part-r-00063  
  inflating: droughtwatch_data/train/part-r-01131  
  inflating: droughtwatch_data/train/part-r-00069  
  inflating: droughtwatch_data/train/part-r-01194  
  inflating: droughtwatch_data/train/part-r-01123  
  inflating: droughtwatch_data/train/part-r-00021  
  inflating: droughtwatch_data/train/part-r-00044  
  inflating: droughtwatch_data/train/part-r-01191  
  inflating: droughtwatch_data/train/part-r-01077  
  inflating: droughtwatch_data/train/part-r-01122  
  inflating: droughtwatch_data/train/part-r-00081  
  inflating: droughtwatch_data/train/part-r-00039  
  inflating: droughtwatch_data/train/part-r-01013  
  inflating: droughtwatch_data/train/part-r-00181  
  inflating: droughtwatch_data/train/part-r-01146  
  inflating: droughtwatch_data/train/part-r-01044  
  inflating: droughtwatch_data/train/part-r-00068  
  inflating: droughtwatch_data/train/part-r-00188  
  inflating: droughtwatch_data/train/part-r-01070  
  inflating: droughtwatch_data/train/part-r-01021  
  inflating: droughtwatch_data/train/part-r-01042  
  inflating: droughtwatch_data/train/part-r-01047  
  inflating: droughtwatch_data/train/part-r-00107  
  inflating: droughtwatch_data/train/part-r-00169  
  inflating: droughtwatch_data/train/part-r-00056  
  inflating: droughtwatch_data/train/part-r-01190  
  inflating: droughtwatch_data/train/part-r-00072  
  inflating: droughtwatch_data/train/part-r-01088  
  inflating: droughtwatch_data/train/part-r-01186  
  inflating: droughtwatch_data/train/part-r-01130  
  inflating: droughtwatch_data/train/part-r-00127  
  inflating: droughtwatch_data/train/part-r-00159  
  inflating: droughtwatch_data/train/part-r-00019  
  inflating: droughtwatch_data/train/part-r-01111  
  inflating: droughtwatch_data/train/part-r-01016  
  inflating: droughtwatch_data/train/part-r-00036  
  inflating: droughtwatch_data/train/part-r-01105  
  inflating: droughtwatch_data/train/part-r-01031  
  inflating: droughtwatch_data/train/part-r-00163  
  inflating: droughtwatch_data/train/part-r-01129  
  inflating: droughtwatch_data/train/part-r-00092  
  inflating: droughtwatch_data/train/part-r-00117  
  inflating: droughtwatch_data/train/part-r-00049  
  inflating: droughtwatch_data/train/part-r-00170  
  inflating: droughtwatch_data/train/part-r-00022

Visualizing data

In [ ]:
import os
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf

The dataset is in the form of TFRecordDataset. In a TFRecord dataset, data is serialized into a binary format, which reduces storage space and speeds up data reading and processing. Each record in the dataset contains a serialized example, which typically includes features and labels for machine learning tasks.

In [ ]:
import os
import tensorflow as tf

# Define a lambda function to generate a list of file paths
dirlist = lambda di: [os.path.join(di, file) for file in os.listdir(di) if 'part-' in file]

# Get a list of training file paths
training_files = dirlist('data/val/')

# Define a function to parse TFRecord data
def parse_visual(data):
    # Create a TFRecordDataset from the input data file
    dataset = tf.data.TFRecordDataset(data)

    # Define the features expected in the TFRecord examples
    features = {
        'B2': tf.io.FixedLenFeature([], tf.string),  # Blue band data
        'B3': tf.io.FixedLenFeature([], tf.string),  # Green band data
        'B4': tf.io.FixedLenFeature([], tf.string),  # Red band data
        'label': tf.io.FixedLenFeature([], tf.int64),  # Label or target value
    }

    # Parse each TFRecord example in the dataset using the defined features
    parsed_examples = [tf.io.parse_single_example(data, features) for data in dataset]
    return parsed_examples

# Parse TFRecord examples from the first training file
parsed_examples = parse_visual(training_files[0])
In [ ]:
def get_img_from_example(parsed_example, intensify=True):
    # Initialize an empty RGB array with dimensions 65x65x3
    rgbArray = np.zeros((65,65,3), 'uint8')

    # Iterate over each band (B4, B3, B2)
    for i, band in enumerate(['B4', 'B3', 'B2']):
        # Extract band data from the parsed example and convert it to numpy array
        band_data = np.frombuffer(parsed_example[band].numpy(), dtype=np.uint8)
        # Reshape the band data to match the dimensions of a 65x65 image
        band_data = band_data.reshape(65, 65)


        if intensify:
            band_data = band_data/np.max(band_data)*255
        else:
            band_data = band_data*255

        # Assign the band data to the corresponding channel in the RGB array
        rgbArray[..., i] = band_data

    # Extract the label from the parsed example and convert it to numpy array
    label = tf.cast(parsed_example['label'], tf.int32).numpy()

    return rgbArray, label
In [ ]:
fig=plt.figure(figsize=(20, 30), dpi= 80, facecolor='w', edgecolor='k')
for i in range(1,26):
    plt.subplot(5, 5, i)
    img, label = img, label = get_img_from_example(parsed_examples[i+7])
    plt.imshow(img).axes.get_xaxis().set_visible(False)
    plt.imshow(img).axes.get_yaxis().set_visible(False)
    plt.title(str(label))

fig.show()
No description has been provided for this image

Defining constants

In [ ]:
# Total size of the dataset
TOTAL_TRAIN = 86317
TOTAL_VAL = 10778

# Sample size of dataset for training purpose
SIZE= 0.1 # modify this only
SIZE_TRAIN = int(TOTAL_TRAIN*SIZE)
SIZE_VAL = int(TOTAL_VAL)

# Parameters of the data (do not change)
IMG_DIM = 65
NUM_CLASSES = 4

Defining features

The satellite imagery data contains different spectral bands. These bands represent different wavelengths of light captured by the satellite sensor, each providing unique information about the Earth's surface.

B1 30 meters 0.43 - 0.45 µm Coastal aerosol

B2 30 meters 0.45 - 0.51 µm Blue

B3 30 meters 0.53 - 0.59 µm Green

B4 30 meters 0.64 - 0.67 µm Red

B5 30 meters 0.85 - 0.88 µm Near infrared

B6 30 meters 1.57 - 1.65 µm Shortwave infrared 1

B7 30 meters 2.11 - 2.29 µm Shortwave infrared 2

B8 15 meters 0.52 - 0.90 µm Band 8 Panchromatic

B9 15 meters 1.36 - 1.38 µm Cirrus

B10 30 meters 10.60 - 11.19 µm Thermal infrared 1, resampled from 100m to 30m

B11 30 meters 11.50 - 12.51 µm Thermal infrared 2, resampled from 100m to 30m

We create a dictionary 'features' which defines the structure of the data stored in TFRecord format. The need to do this arises because:

Data Structure Definition: The TFRecord format is a binary storage format used in TensorFlow for efficiently handling large datasets. However, it's a generic format that doesn't inherently understand the structure of the data it contains. By creating this dictionary, we explicitly define the structure of the data that will be stored in TFRecord files. Each key-value pair in the dictionary represents a feature name and its corresponding properties, such as data type and shape.

Parsing: When reading data from TFRecord files during training or inference, TensorFlow needs to know how to interpret the binary data stored in each example. The dictionary serves as a blueprint for parsing the TFRecord examples correctly. TensorFlow uses this dictionary to decode the binary data into tensors that can be fed into the machine learning model.

Consistency and Compatibility: Defining the data structure upfront ensures consistency and compatibility between the data stored in TFRecord files and the model's input requirements. It ensures that the features expected by the model match the features stored in the TFRecord files, preventing data parsing errors and mismatches during training or inference.

Integration with TensorFlow APIs: TensorFlow provides high-level APIs for working with TFRecord datasets, such as tf.data.TFRecordDataset and tf.io.parse_single_example. These APIs rely on the feature dictionary to understand how to read and parse the data correctly.

In [ ]:
features = {
  'B1': tf.io.FixedLenFeature([], tf.string),
  'B2': tf.io.FixedLenFeature([], tf.string),
  'B3': tf.io.FixedLenFeature([], tf.string),
  'B4': tf.io.FixedLenFeature([], tf.string),
  'B5': tf.io.FixedLenFeature([], tf.string),
  'B6': tf.io.FixedLenFeature([], tf.string),
  'B7': tf.io.FixedLenFeature([], tf.string),
  'B8': tf.io.FixedLenFeature([], tf.string),
  'B9': tf.io.FixedLenFeature([], tf.string),
  'B10': tf.io.FixedLenFeature([], tf.string),
  'B11': tf.io.FixedLenFeature([], tf.string),
  'label': tf.io.FixedLenFeature([], tf.int64),
}

Extracting training and validation data

In [ ]:
import os
import tensorflow as tf

def get_data(train_data_size, val_data_size, local=True):

    def load_data_local(data_path):
        train = file_list_from_folder("train", data_path)
        val = file_list_from_folder("val", data_path)
        return train, val

    def file_list_from_folder(folder, data_path):
        folderpath = os.path.join(data_path, folder)
        filelist = []
        for filename in os.listdir(folderpath):
            if filename.startswith('part-') and not filename.endswith('gstmp'):
                filelist.append(os.path.join(folderpath, filename))
        return filelist

    def parse_tfrecords(filelist, batch_size, buffer_size, include_viz=False):
        # try a subset of possible bands
        def _parse_(serialized_example, keylist=['B1', 'B4', 'B3', 'B2', 'B5', 'B6', 'B7', 'B8', 'B9', 'B10', 'B11']):
            example = tf.io.parse_single_example(serialized_example, features)

            def getband(example_key):
                img = tf.io.decode_raw(example_key, tf.uint8)
                return tf.reshape(img[:IMG_DIM**2], shape=(IMG_DIM, IMG_DIM, 1))

            bandlist = [getband(example[key]) for key in keylist]
            # combine bands into tensor
            image = tf.concat(bandlist, -1)

            # one-hot encode ground truth labels
            label = tf.cast(example['label'], tf.int32)
            label = tf.one_hot(label, NUM_CLASSES)

            return {'image': image}, label

        tfrecord_dataset = tf.data.TFRecordDataset(filelist)
        tfrecord_dataset = tfrecord_dataset.map(lambda x:_parse_(x)).shuffle(buffer_size).repeat(-1).batch(batch_size)
        tfrecord_iterator = iter(tfrecord_dataset)
        image, label = tfrecord_iterator.get_next()
        return image, label

    if local:
        data_path = "/content/data/"
        train_tfrecords, val_tfrecords = load_data_local(data_path)

    X_train, y_train = parse_tfrecords(train_tfrecords, train_data_size, train_data_size)
    X_val, y_val = parse_tfrecords(val_tfrecords, val_data_size, val_data_size)

    return X_train, X_val, y_train, y_val



X_train_total, X_val_total, y_train_total, y_val_total = get_data(SIZE_TRAIN, SIZE_VAL, local=True)

The holdout fucntion implements a holdout strategy for splitting a dataset into training, validation, and test sets.

In [ ]:
def holdout(X_train_total, X_val_total, y_train_total, y_val_total, proportion=(2/3)):
    '''Hold out function'''

    k = int(proportion * SIZE_TRAIN) # Modify this only

    X_train, y_train = X_train_total["image"][:k], y_train_total[:k]
    X_val, y_val = X_train_total["image"][k:], y_train_total[k:]
    X_test, y_test = X_val_total["image"], y_val_total

    return X_train, y_train, X_val, y_val, X_test, y_test
X_train, y_train, X_val, y_val, X_test, y_test = holdout(X_train_total, X_val_total, y_train_total, y_val_total)

Cleaning the data

In [ ]:
import numpy as np

def clean_data(X,y):
    ''' Delete empty images (std < 10) '''
    def find_empty_images(X):
        empty_images = []
        X = np.array(X)
        for i in range(X.shape[0]):
            if X[i].std() < 10:
                empty_images.append(i)
        return empty_images
    X = np.array(X)
    y = np.array(y)
    empty_imgs = find_empty_images(X)
    new_index = [i for i in range(X.shape[0]) if i not in empty_imgs]
    X = np.take(X,new_index,axis=0)
    y = np.take(y,new_index,axis=0)
    return X,y

X_train, y_train = clean_data(X_train, y_train)
X_val, y_val = clean_data(X_val, y_val)
X_test, y_test = clean_data(X_test, y_test)

when working with multi-channel image data, such as satellite imagery, selecting specific channels can help focus the model's attention on relevant information, reduce computational complexity, and improve model performance by discarding irrelevant or redundant channels. Here we define a function to select specific channels from a dataset of images and returns a modified dataset containing only the selected channels.

In [ ]:
def dataset_select_channels(train_images,list_of_channels):
    ''' Input a dataset and a list of channels as a list:
    Example: dataset_select_channels(train_images,['B4','B3','B2']) to convert
    images to RGB.
    '''
    channels_index = [features_list.index(i) for i in list_of_channels]
    data = np.array(train_images)
    return data[:,:,:,channels_index]

features_list = [ 'B1', 'B4', 'B3', 'B2', 'B5', 'B6', 'B7', 'B8', 'B9', 'B10', 'B11']
 # Select features
list_of_channels = ['B7', 'B6', 'B5'] # Modify this for each model, full_list_of_channels = ['B1','B4', 'B3', 'B2', 'B5', 'B6', 'B7','B8','B9','B10','B11']
X_train = dataset_select_channels(X_train, list_of_channels)
X_val = dataset_select_channels(X_val, list_of_channels)
X_test = dataset_select_channels(X_test, list_of_channels)

In this project we use efficientNet model for classification of satellite images. EfficientNet introduces a novel compound scaling method that uniformly scales the network's depth, width, and resolution in a principled way. This allows EfficientNet to efficiently balance model complexity and computational resources, resulting in better performance across different scales.

image.png

In [ ]:
import os
import sys
import argparse
import math
import pandas as pd
import numpy as np

import tensorflow.compat.v1 as tf
from tensorflow.keras import optimizers
from tensorflow.keras import layers, initializers
from tensorflow.keras import models
from tensorflow.keras.callbacks import EarlyStopping
from tensorflow.keras.applications.vgg16 import VGG16
from tensorflow.keras.applications.vgg16 import preprocess_input
from tensorflow.keras.layers.experimental.preprocessing import Resizing
from tensorflow.keras.applications import EfficientNetB3

from termcolor import colored
from google.cloud import storage
from keras.models import model_from_json

def efficientnet_model():
    ''' Transfer learning model that takes X_train with ['B7','B6','B5']'''

    IMG_SIZE = 65
    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))

    x = inputs
    x = Resizing(65, 65)(x)

    activationnetB3 = EfficientNetB3(include_top=False, weights = "imagenet")(x)

    outputsflatten = layers.Flatten()(activationnetB3)

    outputsdense1 = layers.Dense(64, activation = "relu")(outputsflatten)
    outputsdense2 = layers.Dense(64, activation = "relu")(outputsdense1)

    outputsdense3 = layers.Dense(4, activation = "softmax")(outputsdense2)

    model = tf.keras.Model(inputs, outputsdense3)

    model.compile(optimizer="adam",
                loss="categorical_crossentropy",
                metrics=["accuracy"])

    return model
In [ ]:
def train_efficient_net(X_train, X_val, y_train, y_val):
    '''Function to fit Efficient Net model'''

    es = EarlyStopping(monitor='val_loss', patience=20, verbose=1, restore_best_weights=True)

    datagen = tf.keras.preprocessing.image.ImageDataGenerator()
    datagen.fit(X_train)

    X_val = Resizing(65, 65, interpolation="bilinear")(X_val)

    history = model.fit(datagen.flow(X_train, y_train, batch_size=16),
                        epochs=50,
                        validation_data = (X_val, y_val),
                        verbose = 1,
                        callbacks=[es])

    return history
In [ ]:
history =train_efficient_net(X_train, X_val, y_train, y_val)

Epoch 1/50
356/356 [==============================] - 127s 165ms/step - loss: 1.0772 - accuracy: 0.5864 - val_loss: 1.3423 - val_accuracy: 0.5862
Epoch 2/50
356/356 [==============================] - 53s 150ms/step - loss: 0.9970 - accuracy: 0.6083 - val_loss: 2.0164 - val_accuracy: 0.5967
Epoch 3/50
356/356 [==============================] - 64s 180ms/step - loss: 0.9669 - accuracy: 0.6080 - val_loss: 0.9412 - val_accuracy: 0.6030
Epoch 4/50
356/356 [==============================] - 62s 174ms/step - loss: 0.9691 - accuracy: 0.6018 - val_loss: 0.9759 - val_accuracy: 0.6048
Epoch 5/50
356/356 [==============================] - 61s 171ms/step - loss: 0.9953 - accuracy: 0.6074 - val_loss: 1.1164 - val_accuracy: 0.6048
Epoch 6/50
356/356 [==============================] - 60s 167ms/step - loss: 0.9601 - accuracy: 0.6031 - val_loss: 1.0989 - val_accuracy: 0.6048
Epoch 7/50
356/356 [==============================] - 58s 163ms/step - loss: 0.9383 - accuracy: 0.6088 - val_loss: 0.9617 - val_accuracy: 0.5988
Epoch 8/50
356/356 [==============================] - 60s 168ms/step - loss: 0.9211 - accuracy: 0.6183 - val_loss: 0.9464 - val_accuracy: 0.6205
Epoch 9/50
356/356 [==============================] - 63s 179ms/step - loss: 0.8892 - accuracy: 0.6441 - val_loss: 0.9397 - val_accuracy: 0.6398
Epoch 10/50
356/356 [==============================] - 62s 175ms/step - loss: 0.8634 - accuracy: 0.6538 - val_loss: 0.8646 - val_accuracy: 0.6493
Epoch 11/50
356/356 [==============================] - 59s 167ms/step - loss: 0.8486 - accuracy: 0.6547 - val_loss: 0.9940 - val_accuracy: 0.6289
Epoch 12/50
356/356 [==============================] - 65s 182ms/step - loss: 0.8432 - accuracy: 0.6634 - val_loss: 0.9236 - val_accuracy: 0.6240
Epoch 13/50
356/356 [==============================] - 64s 181ms/step - loss: 0.8334 - accuracy: 0.6608 - val_loss: 0.8994 - val_accuracy: 0.6549
Epoch 14/50
356/356 [==============================] - 62s 173ms/step - loss: 0.7738 - accuracy: 0.6833 - val_loss: 1.0798 - val_accuracy: 0.6430
Epoch 15/50
356/356 [==============================] - 55s 153ms/step - loss: 0.7767 - accuracy: 0.6838 - val_loss: 1.0439 - val_accuracy: 0.6356
Epoch 16/50
356/356 [==============================] - 60s 168ms/step - loss: 0.7524 - accuracy: 0.6961 - val_loss: 1.0306 - val_accuracy: 0.6692
Epoch 17/50
356/356 [==============================] - 61s 172ms/step - loss: 0.7062 - accuracy: 0.7117 - val_loss: 1.6739 - val_accuracy: 0.6282
Epoch 18/50
356/356 [==============================] - 60s 169ms/step - loss: 0.7049 - accuracy: 0.7205 - val_loss: 1.7295 - val_accuracy: 0.6310
Epoch 19/50
356/356 [==============================] - 57s 160ms/step - loss: 0.7070 - accuracy: 0.7212 - val_loss: 0.9054 - val_accuracy: 0.6559
Epoch 20/50
356/356 [==============================] - 63s 177ms/step - loss: 0.6562 - accuracy: 0.7425 - val_loss: 3.2931 - val_accuracy: 0.6135
Epoch 21/50
356/356 [==============================] - 63s 177ms/step - loss: 0.6672 - accuracy: 0.7363 - val_loss: 1.0890 - val_accuracy: 0.6496
Epoch 22/50
356/356 [==============================] - 61s 171ms/step - loss: 0.6414 - accuracy: 0.7519 - val_loss: 1.2751 - val_accuracy: 0.6542
Epoch 23/50
356/356 [==============================] - 60s 169ms/step - loss: 0.5726 - accuracy: 0.7791 - val_loss: 1.0321 - val_accuracy: 0.6433
Epoch 24/50
328/356 [==========================>...] - ETA: 4s - loss: 0.5317 - accuracy: 0.8028

Accuracy achieved in the classification task using effcientNet model:

In [ ]:
results = model.evaluate(X_test,y_test,verbose=1)
print(f'The accuracy of the model is:{results[1]}')
print(results)

334/334 [==============================] - 175s 513ms/step - loss: 1.2502 - accuracy: 0.4534
The accuracy of the model is:0.45342081785202026
[1.2502360343933105, 0.45342081785202026]

References:

  • Tan, M., & Le, Q. V. (2020). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv preprint arXiv:1905.11946.
  • Hobbs, A., & Svetlichnaya, S. (2020). Satellite-based Prediction of Forage Conditions for Livestock in Northern Kenya. arXiv preprint arXiv:2004.04081.
  • https://github.com/wandb/droughtwatch